The neural dynamics of feature-based attention Ruobing Xia B.S. Physics, Nanjing University, 2011 A thesis dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the department of Neuroscience at Brown University Providence, Rhode Island May 2019 © Copyright 2019 by Ruobing Xia This dissertation by Ruobing Xia is accepted in its present form by the Department of Neuroscience as satisfying the dissertation requirements for the degree of Doctor of Philosophy Date: David Sheinberg, PhD., Advisor Department of Neuroscience, Brown University Recommended to the Graduate Council Date: Christopher Moore, Ph.D., Reader Department of Neuroscience, Brown University Date: Thomas Serre, Ph.D., Reader Department of Cognitive, Linguistic and Psychological Sciences, Brown University Date: Behrad Noudoost, Ph.D., Reader Department of Ophthalmology/Visual Sciences, University of Utah School of Medicine Approved by the Graduate Council Date: Andrew G. Campbell, Ph.D., Dean of Graduate School, Brown University iii Curriculum Vitae Ruobing Xia ( ) Email: ruobingxia42@gmail.com Education 2012-now Ph.D. student in Neuroscience, Brown University Neurophysiology, psychophysics and modeling of the feature-based attention in non-human primates and humans. Advisor: David Sheinberg. 2007- B.S. in Physics, Nanjing University, China 2011 Optics & biophysics. Chief editor for a student-led physics magazine. Experience 2013 Teaching Assistant, Experimental Neurobiology Course, Brown University Guiding neurophysiological lab sessions 2010- Research Assistant, Institute of Neurosciences, Chinese Academy of Sciences 2012 Psychophysics and modeling of the visuomotor transformation Academic Projects • [Leading] Studying the network mechanism of feature-based attention in monkey’s area V4 using multielectrode recordings and synchrony analyses. • [Leading] Investigating the behavioral and neural dynamics of visual attention in human by combining psychophysical tests and EEG recordings. • [Leading] Chronic multi-channel optogenetic stimulations with recordings in NHP. • [Collaborating with Serre lab] Building a large neurophysiological dataset in order to decipher the neural code of vision using deep neural networks. • [Collaborating] Studying the interaction between feedforward and feedback visual processing with simultaneous recordings in IT and V4. • [Pre-doctoral] Building a continuous attractor network model to explain the saccadic and microsaccadic eye movement generation pattern. • [Pre-doctoral] Psychophysics studies on the spatial coordinate and time course of visuomotor transformation. Non-academic Projects • [Team] Data-driven analysis of the Society of Neuroscience (SfN) Annual Meeting: topics, trends, and collaborations. http://sfn.metaneuro.com • [Collaborating] Building an open-source Python-based toolbox for electrophysiological data analyses. https://github.com/SummitKwan/PyNeuroSG iv Publications A multilayered story of memory retrieval. Ruobing Xia, Shaobo Guan & David Sheinberg. Neuron (Preview article), 2015. Long term retention of saccadic adaptation is induced by a dark environmental context. Jing Wang, Ruobing Xia, Mingsha Zhang & Yujun Pan. Brain Research, 2012. Covert attention regulates saccadic reaction time by routing between different visual- oculomotor pathways. Shaobo Guan, Yu Liu, Ruobing Xia & Mingsha Zhang. Journal of Neurophysiology, 2012. Abstracts and presentations Dynamic lateral interactions in monkey area V4. Ruobing Xia, Shaobo Guan & David Sheinberg. Society for Neuroscience annual meeting, 2017. Data-driven analysis of Society for Neuroscience (SfN) abstracts. Shaobo Guan, Ruobing, Mitchell Wortsman, Justin Young. Brown Digital Repository, 2016. A unified network model for microsaccade and macrosaccade generation. Ruobing Xia. Vision Science Society annual meeting, 2015. Feature-based attention regulates long-range neural interactions in monkey area V4. Society for Neuroscience annual meeting, 2015. Academic activities 2017-now EPSCoR Attention Consortium trainee. 2013-2017 Neural algorithm journal club. Statistical neuroscience lunch meetings. Neural coding journal club. 2013 Brown-MBL NeuroPracticum program. 2012 Computational and Cognitive Neurosciences Summer School (Cold Spring Harbor Asia), Tsinghua University, China. 2011-2012 Computational neuroscience study group. Skills • [Neuroscience] Vision research, invasive electrophysiology, optogenetics, EEG, psychophysics, behavioral training, animal surgeries, modeling. • [Statistics] Signal processing & time series analysis, machine learning, hypothesis testing, data visualization, computational vision. • [Programming] Proficient in Python, MATLAB; basic knowledge in tensorflow, d3.js, SQL, openGL. v Preface and acknowledgement How long can a day be? When you stand next to a monkey, starring at their eyes, listening to the sound of spikes like clock ticks, no drinking and no eating for 8 hours, and repeat it again and again, you know the answer. How dark can a dream be? When you are awakened every night from nightmares where the recording implant is broken, where the chamber gets infection, where the monkey dies when you are about to collect data, or something else that is like the end of world, you know the answer. How lonely can a soul be? When you talk to people about your thoughts and they tell you that makes no sense, when you mention your obstacles and they say that is your own problem, when you are stuck with some real troubles at night and you find that you and your monkey are the only creatures left in the lab, when this furry partner is also leaving you, forever, and you cannot even hold its hand, you know the answer. How warm can words be? When you hear “I know your frustration”, when you hear “whenever you need a friend”, when you hear “there must be a solution”, when you hear “I really enjoyed participating in your study”, when you hear “you are the first person to get this data in the world”, when you hear “this is in good shape”, when you hear “you become much stronger”, you know the answer. How unconditional can the love for science be? When you have seen the starry sky of spike rasters for the first time, when you have learned how to think deeply into the brain, when you have spent 7 years working with a bunch of people with the same faith, when you have gone through all this toughness, you know that you will respect those determined scientists and appreciate their findings so much more, and you will embrace the science all your life . Along the journey, I would like to express my sincere appreciation to my advisor Dr. David Sheinberg for his tremendous support on my project. I was given unlimited freedom when choosing my project, and David has been totally supportive on what I decided to do. His advices every time when I was confused, his help on the experimental system, and his patience and encouragement when things do not come out right really mean a lot to me. I would also like to extend my gratitude to the two other wonderful committee members Dr. Thomas Serre and Dr. Christopher Moore, who have been giving me extremely valuable vi guidance regarding my thesis project and my scientific thinking in general. Importantly, so many thanks to Dr. Behrad Noudoost for kindly willing to travel from Utah to attend my defense as an outside reader. It is my greatest honor. In addition, I wish to thank all my lab colleagues, without whom the projects would not have been finished eventually. They are my mentors, my helpers, my subjects, my critics, and my nicest friends. Special thanks to Diana Burk, for her unbelievably meaningful support whenever I am down (or up) and in need of someone. I also send my thanks to John Ghenne and the Animal Care Facility staff, as they have always been trying their very best to keep the animals happy and healthy. Outside the lab, the neuroscience graduate program is like a large family, and many more department members have helped me a lot as well. Specially I would like to express my thanks to Dr. Gilab Barnea, whose class totally changed my opinion about the molecular neuroscience and brought self-confidence back to me; to Dr. David Berson, with whom I learned what a true scientist looks like; and to Dr. Wilson Truccolo, who is always kindly willing to discuss and answer any questions as precisely as in a class. As to the EEG project, I am extremely grateful to Jerome Sanes, Michael Worden, and Christopher Black for their generous help on building the setup at the minimum cost. Moreover, let me express my deepest gratitude to my parents and my significant other Shaobo Guan. These people care about me more than themselves. Words cannot show how much I am grateful, but thank you for having been standing by my side all the time, thank you for showing me who I am when I get lost, and thank you for making me see the brightness of life and believe in myself. Last, I want to commemorate my baby monkey, Dexter, who accompanied me for 5 years, worked with me almost every single day, tried every version of my experiments and helped me out with everything, who was patient, sweet and considerate, who sadly died due to an unpredictable medical condition. Thank you Dexter, and thanks to all the lab animals for their invaluable contribution to science. vii Table of Contents Abstract ........................................................................................................................................ 1 Chapter 1 Attention: what, where, and how .................................................................. 3 1.1 Attention, perception and action .................................................................................... 3 1.2 The neural substrates of visual attention ................................................................... 6 1.2.1 Frontal-Parietal system: a source of top-down attentional signals...........................6 1.2.2 The attention network ................................................................................................................. 8 1.2.3 Intermediate visual area V4, a center for bottom-up integration and attentional modulation..................................................................................................................................................... 10 1.3 The neural mechanisms of attentional modulation ............................................. 13 1.3.1 Attentional effects on single-neuronal activities............................................................ 13 1.3.2 Recurrent network models of attention ............................................................................ 15 1.3.3 Attentional regulation of neural synchrony ..................................................................... 18 Chapter 2 The network mechanism of feature-based attention in V4 .............. 22 2.1 Motivation ............................................................................................................................ 22 2.2 Results ................................................................................................................................... 24 2.2.1 Experiment design ...................................................................................................................... 24 2.2.2 Characterizing the feature tuning types and receptive fields ................................... 27 2.2.3 Pairwise synchrony depends on the feature domain relationship ......................... 30 2.2.4 Pairwise synchrony depends on the feature preference relationship .................. 32 2.2.5 Single neuron activity is modulated by feature-based attention ............................ 33 2.2.6 Single-channel LFP spectrum is modulated by feature-based attention ............. 35 viii 2.2.7 Pairwise synchrony is modulated by feature-based attention................................. 36 2.2.8 Cross-frequency phase-amplitude coupling is modulated by feature-based attention.......................................................................................................................................................... 40 2.2.9 Synchrony retains similar patterns regardless of the receptive field distance . 42 2.3 Discussion ............................................................................................................................ 45 2.3.1 Summary and significance ....................................................................................................... 45 2.3.2 Common input or recurrent processing? .......................................................................... 48 2.3.3 What do different frequency bands represent? .............................................................. 52 2.3.4 Potential roles of the tuning- and attention-dependent gamma synchrony....... 56 2.3.5 The effect of working memory............................................................................................... 59 2.3.6 Obstacles and lessons ................................................................................................................ 61 2.4 Method .................................................................................................................................. 70 2.4.1 Animal manipulation ................................................................................................................. 70 2.4.2 Behavioral paradigm.................................................................................................................. 71 2.4.3 Visual stimuli................................................................................................................................. 72 2.4.4 Data acquisition and processing ........................................................................................... 72 2.4.5 Feature tuning and receptive field measurement ......................................................... 74 2.4.6 Pairwise synchrony analysis .................................................................................................. 75 Chapter 3 The neural and behavioral dynamics of feature-based attention .. 78 3.1 Motivation ............................................................................................................................ 78 3.2 Results ................................................................................................................................... 80 3.2.1 Experiment design ...................................................................................................................... 80 3.2.2 The spatial-temporal structure of EEG response ........................................................... 82 3.2.3 Detection performance depends on the peri-stimulus alpha phase ...................... 83 3.2.4 The phase dependency is suppressed by feature-based attention......................... 85 ix 3.2.5 Exploring the functional connectivity................................................................................. 87 3.3 Discussion ............................................................................................................................ 89 3.4 Method .................................................................................................................................. 92 3.4.1 Participants .................................................................................................................................... 92 3.4.2 Behavioral Paradigm ................................................................................................................. 93 3.4.3 Visual stimuli................................................................................................................................. 93 3.4.4 Data acquisition and processing ........................................................................................... 94 3.4.5 Data analysis.................................................................................................................................. 95 Chapter 4 Final remarks .................................................................................................. 101 References .............................................................................................................................. 106 x List of Illustrations Figure 1-1. Demonstration of the spatial attention and feature-based attention. _________________________ 4 Figure 1-2 The simplified visual attention network of the primate brain. __________________________________ 9 Figure 1-3. The location of V4 and its connections with other cortical areas. _____________________________ 11 Figure 1-4. Attentional effects on single-neuronal activities. ______________________________________________ 15 Figure 1-5. The recurrent network in the intermediate visual area. _______________________________________ 17 Figure 1-6. The synchrony mechanism of attention modulation. __________________________________________ 21 Figure 2-1. The distributed modular structure in V4 and the hypothesized synchrony pattern. _________ 23 Figure 2-2. The delayed feature match-nonmatch task. ____________________________________________________ 25 Figure 2-3. The feature tuning property of recorded neurons. _____________________________________________ 29 Figure 2-4. The relationship between pairwise synchrony and feature domain. __________________________ 31 Figure 2-5. The relationship between pairwise synchrony and feature selectivity. _______________________ 33 Figure 2-6. The effect of feature-based attention on firing rate. ___________________________________________ 35 Figure 2-7. The effect of feature-based attention on LFP power spectrum. _______________________________ 36 Figure 2-8 The effect of feature-based attention on pairwise synchrony. _________________________________ 39 Figure 2-9. The effect of feature-based attention on phase-amplitude coupling. _________________________ 41 Figure 2-10. The comparison of pairs with closer and farther RF distances. ______________________________ 45 Figure 2-11. A comparison with previous studies. __________________________________________________________ 47 Figure 2-12. Simultaneous optogenetic stimulation and recordings. ______________________________________ 51 Figure 2-13. A conceptual model for exploring the role of synchrony in attention modulation. _________ 58 Figure 2-14. Our original design of task. ____________________________________________________________________ 63 Figure 2-15. A demonstration of the V4 chamber localization. ____________________________________________ 65 Figure 3-1. The attention-related behavioral oscillation. __________________________________________________ 79 Figure 3-2. The task design and behavioral outcome. ______________________________________________________ 81 Figure 3-3. The topographic distribution of post-cue oscillation power changed with time. _____________ 83 xi Figure 3-4. The p values of phase dependency as a function of time and frequency. ______________________ 85 Figure 3-5. The alpha-phase dependency of hit rate under attended and unattended conditions. _______ 86 Figure 3-6. The functional connectivity estimates. _________________________________________________________ 89 Figure 3-7. The channel map of the EEG recording. ________________________________________________________ 95 Figure 3-8. The phase opposition measurement and test. __________________________________________________ 98 List of Tables Table 1. A comparison of neuronal properties in V1, V4 and IT. ............................................................................... 12 Table 2. The criteria of categorizing neuron tuning types for two sets of analyses........................................... 75 xii Abstract Top-down signals such as attention not only modulate an individual neuron’s activity, but also orchestrate the synchrony in sensory areas. For example, spatial attention selectively increases the local gamma coherence for V4 neurons whose receptive fields (RF) overlap with the attended location, presumably enhancing the signal transmission efficacy of a local ensemble. In contrast, feature-based attention requires a larger-scale mechanism, as neurons within the same feature domain are usually distributed throughout the whole retinotopic map. Studies have found that feature-based attention works in such a parallel fashion by regulating neurons according to their feature selectivity, but not RFs. Two important questions emerge: (1) do neurons that share similar feature tuning but are physically far from each other still form a functional ensemble through synchrony, and (2) is this synchrony pattern modulated by feature-based attention? Driven by these questions, my thesis project investigated the network-level dynamics in macaque neocortical area V4 underlying visual processing and feature-based attention modulation. By simultaneously recording the V4 population activity during a feature match task, we found that (1) neural pairs with similar feature tuning properties showed stronger low-frequency synchrony without visual input yet higher gamma synchrony with visual input; (2) feature-based attention modulates the synchrony of neural pairs whose preferred feature matches the attended feature by 1 decreasing the low-frequency synchrony and increasing the gamma synchrony during late visual period; (3) the feature-dependent and attention-modulated synchrony retains similar patterns regardless of the distance between their RF centers. Our results demonstrated a distributed and flexible network mechanism for feature-based attention. To further explore the behavioral effect of neural synchrony and the global brain network pattern underlying feature-based attention, we conducted a complementary project in human subjects using electroencephalogram (EEG) recordings. Our preliminary results indicated that human’s detection performance depends on the phase of alpha-oscillation immediately following stimulus onset, and interestingly, such phase dependency is suppressed by feature-based attention. Additionally, our functional connectivity analysis revealed distinct task-dependent global network patterns with and without consideration of the zero-lag phase synchrony. These findings provide a more global view and additional insights into the neural dynamics of feature-based attention. 2 Chapter 1 Attention: what, where, and how 1.1 Attention, perception and action Standing in front of a wall of bookshelves and searching for a book, you are likely to feel overwhelmed by the large amount of cluttered information rushing into your eyes. That is because recognizing an object using visual input is such a highly demanding task that requires massive computational resources. In order to handle this information overload, the brain has developed a mechanism, known as “selective attention”, which dynamically allocates limited neural resources to predominantly represent and process the most task-relevant information. Therefore, instead of processing everything at the same time, the brain selects potential candidates based on their importance and processes them in a serial manner. Two specific strategies are used to evaluate the priority of the candidates and direct the focus of attention accordingly, either by the location (e.g., the book in search is likely to be located at the top row) or by the feature (e.g., the book has a red spine) of visual objects. These modulations are referred to as spatial attention and feature-based attention, respectively. Attention plays essential roles in our daily tasks, ranging from perceiving and responding to the world in a highly efficient way, to orienting our cognitive activities such as social and memory. 3 Figure 1-1. Demonstration of spatial attention and feature-based attention. In the example of searching for a book in a cluttered bookshelf, different task information leads to different attention modulation. Spatial attention refers to the enhanced perceptual processing of input for a focal location (e.g., the top right corner), and feature-based attention implies a global modulation of representation based on particular features of interest (e.g., the color of the book). How does attention participate in our daily behavior? One metric is the effect that has on our perception and perceptual task performance. As indicated by the term, spatial attention enhances the processing of visual information in a specific part of visual space, i.e., the “attentional spotlight”, while suppressing the rest part of space. As a result, the perceptual sensitivity of stimuli located within the “spotlight” is usually higher than that at unattended locations, as shown by a substantial number of studies (Reviewed by Carrasco, Ling, & Read, 2004). An extreme example of such modulation 4 is the famous “invisible gorilla” test (Simons & Chabris, 1999), where subjects are asked to count the occurrences of ball passing in a video. Focusing on the trajectory of the ball, most subjects totally neglect the “gorilla” walking among the ball players. Similarly, feature-based attention improves the representation of task-relevant feature about visual objects in a scene (e.g., color, shape, texture, moving direction, etc.) regardless of their location, which increases the sensitivity of visual input with the same feature (Bollimunta, Mo, Schroeder, & Ding, 2011; Summerfield & Egner, 2016). Such bias of sensory processing provides a flexible way of allocating our computational resources to prioritize the most useful information. In addition, the selective process of attention has been coupled with the preparation of targeting movements, such as saccades. Under natural visual scenarios, people make saccadic eye movements to explore the environment, and the saccadic targets usually point to the most attention-grabbing parts. The presence of a ‘salience map’ has been hypothesized to describe the topographic distribution of saliency across the visual scene, which is determined by a combination of both bottom-up features and top-down modulation. In the premotor theory of attention, the computation of such attentional saliency maps shares the same neural circuits for sensory-motor integration (Craighero & Rizzolatti, 2005), which implies that the salience map might overlaps with the action priority map (Bisley & Goldberg, 2010). Beyond perceptual tasks and orienting movements, attention is also tightly linked to various cognitive functions. It is thought that attention plays a central role in feature binding, figure-ground segmentation, working memory, long-term memory and 5 perceptual learning, and it is highly intermingled (if not fully overlapped) with the concept of visual consciousness (Ahissar & Hochstein, 2004; Arall, Romeo, & Supèr, 2012; Chun & Turk-browne, 2007; Fougnie, 2008; Treisman, 1998; Tsuchiya & Koch, 2008). Therefore, knowing how attention regulates the neural activity for visual processing can provide important insight to our understanding of the cognitive brain, not limited to the visual system. 1.2 The neural substrates of visual attention 1.2.1 Frontal-Parietal system: a source of top-down attentional signals Where in the brain does the attention signal emerge? As the late stages of the dorsal visual stream (also known as the “where/how” pathway), prefrontal and posterior parietal regions have been thought to play a major role in visual-motor transformation and visual saliency representation, and therefore serve as a source for the selective attention. Lateral intraparietal area (LIP) and the frontal eye fields (FEF) are two well-studied candidate areas. It was reported that, despite some subtle differences in latency and proportion, these two areas both contain a mixture of neurons that represent the importance of visual inputs, the potential movement (e.g., saccade) targets, or both the visual and motor information (Bruce & Goldberg, 1985; Gottlieb, Kusunoki, & Goldberg, 1998). More importantly, the representation of the visuomotor map in these areas does not only depend on their bottom-up features, but also is highly affected by the task-related priority (Bisley & Goldberg, 2010; Ptak, 6 2012) and other higher cognitions, such as working memory and decision making (K. L. Clark, Noudoost, & Moore, 2012; Gold & Shadlen, 2007; Zhang & Barash, 2003). Such a dynamic and integrated representation lays the foundation for attentional selection. Numerous studies have provided causal evidence in support of the idea that the frontal-parietal network serves as the source of attention signal. In human patients, lesions in some of these areas leads to a lack of visual attention to the contralateral visual field, known as the “hemineglect” syndrome (Hillis, Newhart, Heidler, Barker, & Herskovits, 2005). In monkeys, inactivation of either LIP or FEF causes performance deficit in tasks that require attention, whereas micro-stimulation in these areas can induce shift or bias of covert attention, or even directly evokes saccades (Baluch & Itti, 2011). Moreover, it has been shown that the latency for attentional effects in FEF may be even shorter than that in LIP, other frontal areas, and earlier visual regions, suggesting a hierarchical structure for the top-down modulation (Buschman & Miller, 2007; Gregoriou, Gotts, Zhou, & Desimone, 2009). Although LIP and FEF are usually considered as visuospatial areas and most studies focus on their roles in spatial attention, these areas also appear to underlie task dependent feature-based selection. It was found that during a visual search task, the activation of human intraparietal sulcus is modulated by the informative level of feature cues aside from the spatial cue direction (Egner et al., 2008). Recordings in monkey also revealed that FEF neurons are modulated by feature-based attention with a shorter latency than visual cortex (Zhou & Desimone, 2011). A more recent 7 study identified a source region of feature-based attention, a region called the ventral prearcuate (VPA) in the prefrontal cortex, which appears to causally contribute to the feature-based representation in FEF. In summary, the frontal-parietal network acts as the driving source of both spatial and feature-based attention. 1.2.2 The attention network The implementation of attentional selection and modulation requires a whole-brain network. Take the red and blue arrows in Figure 1-2 as a simplified example. The frontal-parietal system integrates bottom-up information and serves as the “top” of the top-down modulation, whose role in part is to determine the selection of targets (in the spatial or feature domain). The “down” end of this modulation signal is the visual system, such as area V4 and middle temporal area (MT), where the input visual information is processed and is affected by the attentional modulation. However, there are many more structures underlying attention processing than this simple bidirectional loop. For instance, the pulvinar nuclei have largely attracted people’s interest recently. As thalamic nuclei, the pulvinar predominately connects reciprocally with cortical regions (as opposed relaying information from the neural periphery) including both the visual areas (e.g., V4) and the frontal-parietal network (Figure 1-2). Electrophysiological and inactivation studies show that this structure plays a critical role in coordinating and maintaining the activation and synchrony patterns across regions according to the attention allocation (Fiebelkorn, Pinsk, & Kastner, 2019; Meldolesi et al., 2012; Zhou, Schafer, & Desimone, 2016). Other brain structures like the superior colliculus (SC), temporal parietal junction (TPJ), lateral 8 prefrontal cortex (LPFC), etc. are also involved. The specific role of each region in the network is not yet fully understood. Figure 1-2 The simplified visual attention network of the primate brain. In the brain diagram, blue arrows represent bottom-up projections, red arrows represent top-down projections, black arrows represent recurrent connections, and gray arrows represent the bidirectional corticothalamic interactions (Miller & Buschman, 2012). 9 1.2.3 Intermediate visual area V4, a center for bottom-up integration and attentional modulation To understand the mechanisms of attentional modulation, many studies have investigated the top-down regulation of the neuronal responses in area V4 of the macaque monkey, an intermediate level of the primate visual system. We will first review the anatomical and functional properties of V4 and discuss why it is an ideal region for studying visual attention. V4 sits in the center of the visual system. Anatomical studies using both anterograde and retrograde tracing methods showed that V4 has bi-directional topographic connections with both lower visual area (e.g., V1, V2) and higher ventral regions (e.g., IT), consistent with its intermediate position in the ventral stream (Figure 1-3) (Ungerleider, Galkin, Desimone, & Gattass, 2008). Interestingly, V4 is also reciprocally connected with many dorsal stream areas, especially LIP and FEF, the putative “top” of the top-down modulation. It is worth noting that only the peripheral field of V4, but not the foveal part, projects to LIP and other parietal areas, suggesting an auxiliary pathway by which visual information of peripheral events can be directly transmitted to dorsal regions for attention allocation; in contrast, the feedback projection from dorsal areas target on the whole visual field of V4, indicating a global effect of top- down regulation (Ungerleider et al., 2008). Therefore, as a hub of the visual system, V4 not only sits in the middle of the ventral pathway, but also builds a bridge between the ventral stream and the dorsal attention system. 10 Figure 1-3. The location of V4 and its connections with other cortical areas. Closed arrows indicate feedforward projections, and open arrows indicate feedback projections (Ungerleider et al., 2008). Functional imaging and electrophysiological recording further revealed that V4 is a mid-tier visual area for feature representation, integration, and transformation (reviewed by Roe et al., 2012). Compared with earlier visual areas (e.g., V1), V4 neurons have larger receptive fields and are tuned to more sophisticated features, such as hue, curvature and 3D surface (Essen, 2005; Li, Liu, Juusola, & Tang, 2014; Pasupathy & Connor, 1999), instead of color opponent, orientation and 2D boundary. Moreover, multiple transformations appear to occur in V4 (Table 1). For example, the color tuning (or receptive field) may shift due to different background colors (or spatial translation), so that a neuron’s response is invariant to context. In the higher area IT, neurons have even larger receptive fields, sparser feature selectivity and greater contextual invariance, indicating that visual information is hierarchically processed through the ventral visual pathway. In other words, as an intermediate 11 stage, V4 has a higher level of information integration than V1 yet a finer spatial resolution than IT, suggesting V4 might be heavily involved in spatial-feature integration, such as binding and segmentation, as well as spatial and feature-based modulation. Properties V1 V4 IT RF size 0.1 – 5 deg 1 - 10 deg 15-30 deg Hue map Color Color opponent Color constancy Curvature Shape Orientation Selective to objects or Some invariance object components Relative disparity Greater invariance Depth Absolute disparity (3D surface) Finer motion Motion Direction Motion-defined shape Attention ~800 ms ~150 ms Similar to V4 (latency) Table 1. A comparison of neuronal properties in V1, V4 and IT. (Buffalo, Fries, Landman, Liang, & Desimone, 2010; Roe et al., 2012) As mentioned before, V4 also receives projections from the dorsal attentional regions. The central position of V4 in the ventral pathway makes it possible that top-down signals are sent to V4 and then propagated throughout the ventral visual system. Evidence comes from studies that record and compare multiple brain areas. Whereas the lower areas V1 and V2 show much weaker attentional modulation, the activity in V4 is dramatically regulated by both spatial and feature-based attentional cues (Roe et al., 2012). More strikingly, a recent study that compared the latency of attention- induced activity enhancement at multiple visual areas suggested that attentional 12 modulation is propagated from V4 to V2 and V1 (Table 1) (Buffalo et al., 2010). These results directly support the idea that V4 plays a central role in the attentional modulation network. 1.3 The neural mechanisms of attentional modulation 1.3.1 Attentional effects on single-neuronal activities In this section, we will use V4 as a well-studied and representative example area to show what is known about the mechanisms of attentional modulation. Traditionally, most studies focused on the modulation of single unit activity, such as gain modulation. It was first found that the response intensity (i.e., firing rate) of a V4 neuron was up-regulated or down-regulated depending on whether a target or a distractor appeared in its receptive field, respectively (McAdams & Maunsell, 1999; Moran & Desimone, 1985). Similarly, feature-based attention induced by a pre- stimulus cue led to an enhancing gain modulation of neuronal response when both the visual stimulus and the cue were consistent with the neuron’s preferred feature, regardless of their spatial selectivity, as shown in Figure 1-4a (Bichot, Rossi, & Desimone, 2005; Zhou & Desimone, 2011). Attention toward a certain feature also caused suppression of the surrounding representation in the feature space (Störmer & Alvarez, 2014). Importantly, the effect of gain modulation became more robust when target and distractor were simultaneously located in the receptive field, 13 indicating that gain modulation is effective especially when multiple stimuli are competing with each other, highlighting the gating role of attention (S. Luck & Chelazzi, 1997). In terms of the form of gain modulation, there has been much debate on whether it is multiplicative gain or additive gain, and whether attention regulates the peak response (response gain) of neurons or only influences their sensitivity to lower contrast stimuli (contrast gain) (Herrmann, Montaser-Kouhsari, Carrasco, & Heeger, 2010; Ling & Carrasco, 2005). Despite the controversy, both spatial attention and feature-based attention can enhance the response strength of neurons that encode the attended object. In addition to gain modulation, it was also found that the tuning curves of V4 neurons shift due to attention. Tuning curve shifts were first observed in a study that measured the spatial tuning of V4 neurons, i.e., receptive fields. In the presence of a spatial cue (e.g., an arrow to the right of fixation spot), the spatial tuning profile of V4 neurons shifted toward the attended direction (e.g., right) (Connor, Gallant, Preddie, & Van Essen, 1996). Similarly, feature-based attention also shifts the feature tuning curves of V4 neurons. In their study, Gallant and colleagues used natural image patches to measure the orientation and spatial frequency tuning of V4 neurons during a match-to-sample task (David, Hayden, Mazer, & Gallant, 2008). Based on their assessments of attentional effects, the baseline gain modulation could account for ~13% of the response variance, while the tuning modulation was responsible for another ~9% of the variance. These results indicated that attention regulates V4 neurons by both enhancing response amplitude and modulating the center of tuning curves (Figure 1-4b). 14 Figure 1-4. Attentional effects on single-neuronal activities. (a) An example of the gain modulation of single neuronal activity in V4, adapted from Zhou & Desimone, 2011. The red curve represents the activity when the attended target appears in the neuron’s receptive field, and blue curve represents the trials when a distractor that does not share the attended feature is in the receptive field. (b) A demonstration of the tuning curve shift during attention modulation, adapted from David et al., 2008. 1.3.2 Recurrent network models of attention Based on the single-neuron physiological findings, a computational model was proposed to explore the network mechanism underlying the parallel attentional effects – gain modulation and tuning curve shift (Compte & Wang, 2006). This two- layer network model has a multiplicative gain factor in its input layer and recurrent connections in its readout layer (Figure 1-5a). The simulation results showed that when an attention “spotlight” was induced in the input layer, each unit in the readout layer reproduced the reshaped tuning curve with both increasing amplitude and its center shifted toward the focus of attention, suggesting that the tuning curve shift 15 could be a side-effect of gain modulation. It worth noting that the recurrent connection pattern of this model plays a critical role in setting up interactions across the visual field; otherwise, an exclusively feedforward system could produce modulation effects only when the attention spot is almost overlapping with the visual stimulus. Similarly, another set of models that explains the attentional effect on two competitive stimuli, i.e., the normalization model, also highlights the role of recurrent connectivity (Reynolds & Heeger, 2009). A simplified version of this model contains a sensory input, an attention field, and a suppressive field that is determined by the product of sensory input and attentional gain, all of which lead to a normalization effect on the biased representation of both stimuli. The suppressive field, as a key part of the normalization model, was implemented with lateral excitatory and inhibitory projections in a neural network model, again, suggesting the potential existence and importance of recurrent connectivity within visual areas. More broadly, recurrent network models have been used to explore other cognitive functions in addition to attention modulation. Examples include transformation- invariant recognition (Rolls, 2012), adaptivity of tuning functions (Schwabe & Obermayer, 2005), figure-ground segmentation (Roelfsema, Lamme, Spekreijse, & Bosch, 2002), working memory (Camperi & Wang, 1998), winner-take-all decision making (Wang, 2002), etc. In these models, recurrent connection patterns are designed, learned or dynamically modulated so as to produce interesting network 16 properties, such as competitions and attractors, which then contribute to the implementation of higher-level cognitive or computational functions. These models expand the conceptual framework from the single-unit level to a network point of view, especially highlighting the potential role of horizontal connections in V4. Anatomical studies have revealed the connectivity pattern in V4 (Figure 1-5b), with both local connections that cluster in modules and larger-range horizontal connections that build links between modules (Yoshioka, Levitt, & Lund, 1992). Little is known, however, about whether these inter-modular connections underlie the functional structure found in imaging studies (e.g., feature-selective modules in V4), what their functional roles are, and how they are dynamically regulated during visual tasks. A comprehensive investigation of the network properties and the lateral interactions in V4 would therefore provide important insights into our understanding of the network mechanisms of attention. Figure 1-5. The recurrent network in the intermediate visual area. (a) The basic structure of a recurrent network for attention regulation (Compte & Wang, 2006). {x1, …, xn} indicate the selectivity of each neuron in a certain feature space. The model assumes that neurons with similar selectivity have stronger connections. 17 (b) The modular connection pattern in V4 revealed by anatomical tracing methods (Yoshioka et al., 1992). 1.3.3 Attentional regulation of neural synchrony The development of multi-channel recordings and analysis techniques allows us to investigate the interactions between population neurons, which opens up new horizons in the understanding of attention. In primates, the most commonly used way of investigating the effective interactions is to look at the temporal correlation of oscillatory population activities, i.e., synchrony. Previous studies have suggested that synchrony could serve as a flexible mechanism of regulating neuronal interactions and enhancing/suppressing the signal transmission efficacy (Figure 1-6a) (Fries, 2005; Womelsdorf & Fries, 2007). Although the emergence of neural synchrony is still not fully understood, it has been used as an indicator of the functional connectivity for both local networks and large-range communications, and recent attention studies have focused on the synchrony modulation from these two perspectives. An increase of local synchrony has been found in many attention studies. For example, in a cued visual search task, feature-based attention enhanced field-spike coherence within the gamma band when the cue was consistent with the neuron’s preferred feature (Bichot et al., 2005). Similarly, when spatial attention was induced, neurons selective to attended locations showed an increased synchrony in gamma band and a reduced synchrony in alpha band, compared with neurons selective to nearby unattended locations (Fries, Reynolds, Rorie, & Desimone, 2001). Furthermore, a 18 recent study suggested that attention regulates neural synchrony in a laminar- specific manner. Increased gamma-band synchronization mainly occurs in the superficial layers of V4, whereas alpha-band synchronization was decreased in the deep layers (Buffalo, Fries, Landman, Buschman, & Desimone, 2011). Such distributions among layers and frequency bands might differentiate between feedforward and feedback activities (van Kerkoerle et al., 2014). It is generally thought that local field potential (LFP) mainly reflects activity within an area of a few hundred micrometers around the electrode tip (Kajikawa & Schroeder, 2011), and functional modules (e.g., orientation domains) in V4 have a similar size (Roe et al., 2012), which suggests that the local synchrony found in V4 might be an indicator of the interactions within modules (the clusters of functional-similar neurons). Taken together, these results point to the possibility that attention regulates the synchronization of neurons at a modular level. On the other hand, with simultaneous recordings at multiple regions, the analysis of inter-regional synchrony has not only revealed the interaction between brain areas, but also determined the direction of information flow. Paired recordings in FEF and V4 revealed that during a covert attention task, the gamma-band synchronization between FEF and V4 was enhanced only when the pair-wise neurons had overlapping receptive fields. Further analysis on the phase of LFP gamma-band oscillation showed that the phase of FEF oscillation precedes the phase of V4 oscillation by 8-13 milliseconds, suggesting that FEF might be driving attentional effects in V4 through direct synaptic connections (Buschman & Miller, 2007; Gregoriou et al., 2009). Regarding possible reciprocal interactions, V4 was found to gamma-synchronize to 19 V1 when the monkey performed a modified match to sample task. The enhanced synchrony to V1 was dependent on task relevance, meaning that only the task- relevant part of V1 was synchronized to V4. Additionally, spiking of V4 neurons precisely matched the LFP in V1, indicating that V4 drove the response of V1 neurons during attentional regulation (Grothe, Neitzel, Mandon, & Kreiter, 2011). Such gamma synchronization between V1 and V4 was found to predict detection performance of the attended target (Rohenkohl, Bosman, & Fries, 2018). These findings suggest again that V4 acts as the intermediate stage in top-down attention processing, defined by the synchronization relationship between regions, and that the modulation of synchrony strength is highly coupled with the attention-related improvement. 20 Figure 1-6. The synchrony mechanism of attention modulation. (a) A demonstration of “communication-through-synchrony” theory (Fries, 2009). (b) A demonstration of the within-region and cross-region synchrony. (c) An example of the local synchrony in V4 modulated by spatial attention from Buffalo et al., 2011. The gamma-band spike-field coherence appeared to be increased by attention in superficial layers (upper plot), whereas in deep layers (lower plot), the low-frequency coherence was decreased by attention. (d) An example of the FEF-V4 synchrony during attention modulation from Gregoriou et al., 2009. The upper plots show the gamma spike-field coherence between FEF and V4 is up-regulated by attention in both directions. The lower 21 plots show the gamma-band Granger causality between the two regions, with a difference in effect latency suggesting that FEF-> V4 drives the interaction. Chapter 2 The network mechanism of feature-based attention in V4 2.1 Motivation In the first chapter, we set our focus on area V4 as a demonstration of how attention regulates visual processing, with a particular emphasis on synchrony as an emerging potential theory for the attentional mechanism. We reviewed the findings on local and global synchrony, i.e., the attention-related local synchronization and the inter- regional communication of attentional signals. Nevertheless, a key question remains: does attention also affect synchrony beyond the local modules within the same visual area? As revealed by physiological and imaging studies, the features of visual objects are represented in V4 in a fairly distributed fashion (Roe et al., 2012). Different feature modules are interleaved, while at the same time forming a loose retinotopic map (Figure 2-1). Such structure provides an efficient route for attentional modulation in the spatial domain, as neurons preferring similar spatial locations are physically close to each other. A spatial-based top-down signal, therefore, only needs to selectively regulate the activity amplitude and synchrony level of neurons locally (Fries et al., 22 2001). When it comes to feature-based attention, however, neurons that represent the attended features are usually modulated in a parallel manner (Bichot et al., 2005). It seems that a reasonable mechanism should bring these distributed modules together to form a more efficient functional ensemble. Figure 2-1. The distributed modular structure in V4 and the hypothesized synchrony pattern. The optical imaging figure is adapted from Roe et al., 2012, where patches with the same color represent the neural modules that have the same feature selectivity type. A recent study provided some clues. By measuring both firing rates and noise correlations, the authors found that feature-based attention showed a larger effect range (i.e., across hemispheres) than spatial attention which only made local influences on a group of spatial specific neurons (Cohen & Maunsell, 2011). Another study in V1 presented a closer investigation of the horizontal synchronization 23 modulated by top-down signal in relation to task demands (Ramalingam, McManus, Li, & Gilbert, 2013). These studies suggest that larger-range cross-modular interactions could be regulated by attention, yet relevant experimental evidence is still lacking. Based on anatomical and physiological findings, we hypothesize that (1) the functional connection (i.e., synchrony) between neural modules not only depends on their physical distances, but also on their feature selectivity; moreover, (2) this large- scale synchrony pattern is transiently modulated by feature-based attention in a way that improves the communication efficacy of attended modules. Studying the dynamical interaction pattern in V4 will not only complete the framework of attention-evoked synchrony at different levels, but also provide key empirical data for computational models, e.g., recurrent neural networks, and thus improve our understanding of the biological as well as the artificial visual system. 2.2 Results 2.2.1 Experiment design The project aimed to investigate the feature tuning-dependent synchrony in V4 and how feature-based attention modulates this synchrony pattern. To begin to tackle these questions, we trained monkeys to perform a delayed feature match-nonmatch task so as to induce the feature-based top-down modulation (see section 3.2.1 for a detailed description). The task required animals to pay attention to the orientation/color feature of a sample stimulus in order to perform the matching task. 24 The spatial layout of the feature sample and the visual probe (as shown in Figure 2-2) was designed to reduce the visual activity or visual adaptation related to the sample (the sample was shown on the opposite direction of recorded neurons’ receptive fields) and diminish the effect of spatial component (the probe was made large enough to cover almost all recorded receptive fields instead of being presented in or out of the receptive field of specific neurons). After training, both monkeys showed a reasonable performance accuracy for the task (67% and 73%, respectively). Figure 2-2. The delayed feature match-nonmatch task. At time 0 of each trial, a feature sample (either a gray grating or a colored round patch) is shown outside the recorded neurons’ receptive fields. Followed by a 0.5 s delay period, a visual probe (a colored grating) is presented to cover all receptive fields. The subject is required to maintain fixation until the fixation spot goes off, and then make a button press to indicate whether the visual probe contains the same feature with the sample. 25 At the same time, we took advantage of a multielectrode recording system (Gray Matter Research SC-32 microdrive) to collect spiking activity and LFP signals simultaneously from multiple sites in V4 when the animal was performing the feature-match task. In this microdrive setup, the physical distance between neighboring electrodes is 1.5 mm, which allows us to focus on the distributed network across a relatively large scale, as opposed to looking at local connectivity within columns. Our analyses on the physiological data mainly focused on two issues: (1) The synchrony pattern of the V4 horizontal network underlying spontaneous (stimulus absent) and visual (stimulus present) conditions. More specifically, once we characterized the tuning properties of each recorded neuron (section 2.2.2), we asked the question whether such synchrony depended on the tuning relationship between a pair of neurons (sections 2.2.3 and 2.2.4). (2) The regulation of the horizontal synchrony by feature-based attention. We first verified the effectiveness of attention modulation in our task at the single neuron/LFP level (sections 2.2.5 and 2.2.6), and then directly compared the pairwise synchrony pattern between attended and unattended conditions (sections 2.2.7 and 2.2.8). Following these analyses, we finally confirmed the synchrony modulation as a global but not local mechanism by comparing the pairs with different receptive field distances (section 2.2.9). It is important to mention that although we included both animals’ data in the following analysis, we were not able to compare the results for each individual, because the data size in monkey X turned out to be limited (discussed in section 26 2.3.5). This made the results that we show in the next sections mainly represent the phenomena in monkey H. 2.2.2 Characterizing the feature tuning types and receptive fields To investigate the feature-dependent neural dynamics, our initial step was to characterize the feature tuning of recorded neurons. Using Cohen’s d test, we first identified 465 visual-responsive cells out of a total number of 636 recorded cells. The feature tuning of each visual cell was obtained by calculating the mean firing rate during the visual period of the delayed feature-match task. Figure 2-3(a) shows the tuning function of two example neurons with strong orientation selectivity and color selectivity, respectively. An overview of all recorded neurons from one example session is shown in Figure 2-3(b). It is clear that neurons exhibit different selectivity patterns: some are merely sensitive to orientation or color, some demonstrate a combination of selectivity, and some are not modulated by either feature domain. Therefore, we performed 1-way ANOVA for each dimension to quantify the level of selectivity. Figure 2-3(c) shows the distribution of F values in the two dimensions. For the sake of simplicity, we separated these neurons into 4 groups using hard boundaries (see section 2.4.5 for details) despite the continuous distribution. We identified 211 orientation-selective neurons, 100 color-selective neurons, 88 neurons tuned to both feature domains, and the rest showing no reliable selectivity. The colored boxes in Figure 2-3(b) that use the same color scheme as in Figure 2-3(c) demonstrate a reasonable separation of selectivity patterns. 27 In the same recording sessions, we also measured the receptive field (RF) of all recorded units and LFPs with a flashing spot paradigm, defined using spike counts and the gamma-band power of LFP, respectively. Figure 2-3(d) shows the gamma- power-defined RFs averaged across 3 consecutive sessions, which reveals a clear retinotopic map. The distribution of the RF center for all recorded neurons is shown in Figure 2-3(e). The result will be used for later analysis on the relationship between synchrony pattern and RF distances. 28 Figure 2-3. The feature tuning property of recorded neurons. (a) The response heatmap of two example neurons sorted by orientation and color conditions, superimposed with the spike density function under each condition. (b) The tuning heatmap of neurons from an example session with their assigned selectivity type using the same color scheme as in (c). The position of heatmaps follow the layout of the electrodes (c) The population distribution of F values computed from one-way ANOVA on both 29 feature dimensions. The color represents the selectivity type assigned to each neuron. (d) The gamma-power-defined RFs averaged from 3 example sessions. Brighter colors represent higher normalized visual-induced gamma power. White cross marks the foveal point (0,0). (e) The distribution of the RF center of all recorded neurons. 2.2.3 Pairwise synchrony depends on the feature domain relationship We then explored the relationship between the synchrony pattern in V4 and the neurons’ feature tuning. The pairwise synchrony was estimated between each spiking unit and the LFP recorded from other electrodes, and we used the spike-field phase consistency (PPC) as the measure due to its robustness to the change of spike counts (Vinck, van Wingerden, Womelsdorf, Fries, & Pennartz, 2010). The tuning relationship of each pair of signals (spike and LFP) was defined based on the tuning of the spiking unit recorded from the same electrode. Figure 2-4(a, b, d, e) shows the average PPC spectrogram of 4 different kinds of pairs: signals within the same selectivity domain (orientation-orientation or color-color), signals from different domains (orientation-color or color-orientation), signals that both show a combination of selectivity (both-both), and signals with one being the “both” type and one from either feature domain. The difference between these groups are shown in Figure 2-4(c, f). It appears that signals with the same selectivity type, compared with those from different groups, show a stronger low-frequency synchrony during the 30 baseline/delay periods and higher gamma-band synchrony during the late visual period. Figure 2-4. The relationship between pairwise synchrony and feature domain. (a) The average spike-field PPC spectrogram calculated from pairs of the same domain type, i.e., orientation-orientation (O-O) and color-color (C-C) pairs. White bars represent different task phases. (b) The average PPC of pairs from different domains, i.e., orientation-color (O-C) and color-orientation (C-O) pairs. (c) The difference between (a) and (b). (d) The average PPC of pairs that are sensitive to both domains, i.e., both-both (B- B) pairs. (e) The average PPC of pairs with one being sensitive to both domains and the other one showing selectivity to one domain. (f) The difference between (d) and (e). 31 2.2.4 Pairwise synchrony depends on the feature preference relationship Moving one step further, we examined whether the pairwise synchrony not only depends on the feature domain, but also relies on the similarity of feature preferences within the same domain. Due to the limited number of neurons in each domain type, we divided the group of neurons that show selectivity to both dimensions and add them into the orientation and color-only groups. Figure 2-5(a-c) compares the orientation-selective pairs with the same preferred orientation (n = 723) versus those with different preferred orientation (n = 1147). Similarly, Figure 2-5(d-e) shows a comparison of color-selective pairs with the same (n=229) and different (n=345) preferred colors. In both cases, the PPC difference demonstrates a qualitatively similar pattern to Figure 2-4, which indicates that signals with similar feature selectivity tend to have stronger low-frequency synchrony without visual input, and are more gamma synchronized during the late visual period. 32 Figure 2-5. The relationship between pairwise synchrony and feature selectivity. (a, d) The average spike-field PPC spectrogram calculated from pairs of the same domain type (orientation and color, respectively) and the same preferred feature in that domain. White bars represent different task phases. (b, e) The average PPC of pairs from the same domain (orientation and color, respectively) but showing different preferred feature in that domain. (c) The difference between (a) and (b). (f) The difference between (d) and (e). 2.2.5 Single neuron activity is modulated by feature-based attention Having investigated the general feature-dependent synchrony pattern in V4, we next focus on the effect of feature-based attention. We first examined the attentional 33 modulation on single neuron activity to confirm that our task paradigm effectively induces top-down modulation in this area. For each neuron, the feature-based attention effect was defined as the difference between the average visual response when the sample contains the preferred feature of this neuron (i.e., attended condition) versus the response when the sample is not preferred (i.e., unattended condition). In addition, we divided the trials by whether the visual probe contains the preferred feature, and thus we compared 4 conditions in total (Figure 2-6(a)). Figure 2-6(b) shows the normalized spike density function averaged across neurons under the 4 conditions. In general, the sample (on the non-preferred side of visual field) did not induce much visual response, and neither did the delay period. Neurons responded dramatically to the visual probe, with a sharp activity peak followed by milder late activities until the stimulus went off. It can be found that the feature selectivity emerges and sustained consistently throughout the visual response. Importantly, only when the visual probe is preferred, the late visual response shows an effect of attention (the red solid curve is higher than the blue solid curve). In contrast, the response to nonpreferred probe does not show such an effect, which suggests that feature-based attention modulation is more effective on the neurons that are more actively involved in the current visual representation. As a result, we will focus on the “visual preferred” trials in our next analysis. 34 Figure 2-6. The effect of feature-based attention on firing rate. (a) The definition of the 2×2 conditions for analysis, i.e., attended/unattended * preferred/nonpreferred. “Attended” condition is defined as when the sample matches the neuron’s preferred feature. “Preferred” refers to whether the visual probe contains the preferred feature of the neuron. (b) The comparison of the population average response rate under the 4 conditions. Shaded area represents the standard error. The black marks represent the significance of Wilcoxon test between attended (preferred) and unattended (preferred) conditions (p<0.01). 2.2.6 Single-channel LFP spectrum is modulated by feature-based attention Following the single-neuron analysis, we moved our focus to the attentional effect on single-channel LFPs. Using the tuning property of the neuron recorded from the same electrode, we defined the attended and unattended conditions for LFP data with the same method as in the last section. Figure 2-7 shows the comparison between the two 35 conditions. As one would expect, the visual stimulation induced a decrease of theta/alpha band power and an increase of gamma-band power. Furthermore, the contrast shows an enhancement of high-gamma power throughout the late visual period and a decrease of the power in lower frequency bands followed by an increase. Although there was not as strong a modulation as in spatial attention studies (which is consistent with what Bichot et al., 2005 reported), the contrast pattern still showed an effect of feature-based attention, and the result also confirmed that it was reliable to use the tuning of the spiking neurons to represent the property of its local LFP. Figure 2-7. The effect of feature-based attention on LFP power spectrum. (a, b) The average power spectrogram under attended condition (a) and unattended condition (b), respectively. (c) The difference between (a) and (b). 2.2.7 Pairwise synchrony is modulated by feature-based attention To analyze pairwise synchrony, we used a similar method to divide and compare the data as presented in the previous section, but the criteria were based on the 36 preference of neuron pairs rather than individual neurons and we only focused on the visual preferred conditions due to the larger effect (as discussed in the last section). In addition, we only considered the orientation domain, because the current analysis (PPC comparison between conditions) required a large data size and our dataset did not provide enough samples of color-selective pairs. For each pair of neurons that were both orientation-selective and preferred the same orientation, we calculated their spike-field PPC under attended (cued with the preferred orientation, Figure 2-8(a)) and unattended (cued with nonpreferred orientation, Figure 2-8(b)) conditions, respectively. By contrasting them, we obtained the average spectrogram representing the attentional effect (Figure 2-8(c)), which looks noisy and appears to lack modulation at the first glance. However, when we further characterized the population distribution of the PPC and performed pairwise statistical tests, we saw significant modulation of feature-based attention on the synchrony pattern. Figure 2-8(d) shows the PPC averaged in multiple frequency bands. Gray areas mark the nonstationary periods around the visual onset, in which the PPC estimate is not considered as reliable. For each time point, we ran the Wilcoxon signed-rank test and the star-marks represent the significance for a consistent attention modulation (gray: p<0.05, black: p<0.01). We found that for pairs that share similar orientation selectivity, (1) attention toward their preferred orientation decreases their theta/alpha PPC during the delay period and decreases the theta PPC with visual input, whereas (2) it increases the gamma PPC in the late visual period. 37 To confirm that the change of PPC really represents synchrony and not simply task related phase locking, we further performed a permutation shuffling within conditions and ran the analysis again. Figure 2-8(e) shows the PPC traces with the shuffling result removed. The pattern of attentional modulation remains despite the reduced overall level of synchrony, indicating that attention regulates the pairwise synchrony. 38 Figure 2-8 The effect of feature-based attention on pairwise synchrony. (a) The average PPC spectrogram in attended trials (feature sample matches the preferred orientation for the neuron pair of interest). (b) The average PPC spectrogram in unattended trials (feature sample does not match the preferred orientation for the neuron pair of interest). (c) The difference between (a) and (b). 39 (d) Population distribution of the PPC in different frequency bands. Red and blue traces represent attended and unattended conditions, respectively. Shaded area marks the standard error. Gray rectangular areas stand for the nonstationary periods with unreliable estimate. Star-marks indicate the significance of pairwise test between attended and unattended conditions, with gray marks representing p<0.05, and black marks representing p<0.01. (e) The PPC results subtracted by the within-condition permutation result. 2.2.8 Cross-frequency phase-amplitude coupling is modulated by feature-based attention In addition to the spike-field synchrony, we also examined the attentional effect on the cross-frequency coupling between recorded signals. It has been proposed that the amplitude of gamma/high gamma activity can be phase-locked to low-frequency (such as theta-band) oscillations within brain areas or between regions, which might serve as a mechanism for facilitating information transfer between nodes within the large-scale network (Canolty & Knight, 2010). In contrast to gamma synchrony that has been shown to increase during attention modulation, spatial attention seems to reduce the theta-gamma phase-amplitude coupling (PAC) in the extrastriate cortices MT and V4 (Esghaei, Daliri, & Treue, 2015; Spyropoulos, Arturo, & Fries, 2018). Consistent with these findings related to spatial attention, our results reveal the same decreased modulation on the PAC by feature-based attention, but mostly across the alpha (8~10 Hz) and not the theta (3~5 Hz) frequencies (Figure 2-9). After correction with the shuffle permutation result, there remains a significant attentional modulation of the alpha-gamma PAC during the late visual period, indicating that the 40 effect is not fully caused by a change of trial-locked activities. Together with the PPC modulation in Section 2.2.7, these results suggest that feature-based attention might adopt a similar network mechanism as spatial attention, i.e., the enhanced gamma synchrony and reduced low-frequency synchrony as well as low-frequency driving on the gamma activity. Figure 2-9. The effect of feature-based attention on phase-amplitude coupling. (a) The PAC between alpha (8~10 Hz) phase and gamma (35~60 Hz) amplitude averaged over all applicable LFP pairs. Shaded area marks the standard error. 41 Black horizontal bars indicate the significance of pairwise test between attended and unattended conditions (p<0.01). (b) The alpha-gamma PAC subtracted by the within-condition permutation result. (c) The PAC between theta (4~6 Hz) phase and gamma (35~60 Hz) amplitude averaged over all applicable LFP pairs. (b) The theta-gamma PAC subtracted by the within-condition permutation result. 2.2.9 Synchrony retains similar patterns regardless of the receptive field distance Despite the large inter-electrode distance, the neurons we recorded and analyzed still have overlapping RFs sometimes, especially for the foveal ones. To confirm that the synchrony pattern and attention modulation described in the previous sections are not a result of spatial-dependent, local interactions, we divided the pairs used in previous analyses into two groups based on their RF center distance (<2 ̊ or >2 ̊). With the two subgroups, we repeated the same analyses as in Figure 2-4, Figure 2-5, Figure 2-8 and Figure 2-9. A comparison of all results (Figure 2-10) shows that although the overall strength of PPC is much higher for closer pairs, the feature-dependent synchrony pattern (the low-frequency synchrony during sample/delay and the gamma during late visual period) and its modulation of feature-based attention (lower theta and higher gamma synchrony as well as lower cross-frequency coupling in the late visual period) remains similar between closer and farther pairs. We therefore conclude that the feature and attention-dependent synchrony is applied to 42 a global V4 network, not only across a large cortical area, but also among neurons with various spatial preference. 43 44 Figure 2-10. The comparison of pairs with closer and farther RF distances. (a) The feature-dependent synchrony (domain- and feature-specific) compared between the two subgroups of signal pairs. (b) The attentional effect on synchrony patterns (gamma synchrony, theta synchrony, and alpha-gamma phase amplitude coupling) compared between the two subgroups of signal pairs. 2.3 Discussion 2.3.1 Summary and significance In summary, our work explored the network-level dynamics in V4 underlying visual processing and feature-based attention modulation. By simultaneously recording the V4 population activity during a feature match task, we found that (1) neural pairs within the same feature selectivity domain or with similar feature preferences showed stronger low-frequency synchrony without visual stimuli yet higher gamma synchrony with visual input; (2) feature-based attention modulates the synchrony of neural pairs whose preferred feature matches the attended feature by decreasing the low-frequency synchrony and increasing the gamma synchrony during late visual period; (3) the feature-dependent and attention-modulated synchrony retain similar patterns regardless of the distance between their RF centers. These results extend our understanding from previous studies and provide new insights in the following ways. 45 First, our study examined the horizontal network for a larger range within V4 whereas most earlier studies focused on the top-down modulation of individual neural activities or local synchronization. Multiple previous studies have attempted to investigate the tuning-dependent functional connectivity in sensory areas and/or its attentional modulation, either by measuring the spike-spike correlation or spike- field coherence (Bichot et al., 2005; Hillis et al., 2005; Oostenveld, Fries, Maris, & Schoffelen, 2011; Ruff & Cohen, 2014). Most of them used multi-electrode systems with a small tip-to-tip spacing (hundreds of microns). Such a distance between recorded signals would lead to RF partially overlapping and yield higher chance of direct connections between individual neurons or those sharing common input projections. Our recordings, in contrast, were performed with quite large inter- electrode distances (>1.5 mm) while maintaining the spatial resolution compared with large-scale extracortical recordings such as electrocorticography (Figure 2-11a). Therefore, we expect that our results represent the dynamics of a wider network that includes multiple feature modules with distinct spatial preference, which was further confirmed in our control analysis based on RF distances. This spatial coverage is crucial for demonstrating the difference between feature-based and spatial attention, i.e., the parallel regulation versus a focal mechanism. Meanwhile, our paradigm also allowed us to mainly focus on the global feature-based modulation. As a comparison, for example, the traditional visual search task usually presents multiple visual items following a feature cue (Figure 2-11b). The neural response thus highly depends on the spatial configuration of the stimuli, and the attention effect can be seen as a combination of the non-spatial cueing effect and a 46 spatial/object-based component. However, our task was designed to reduce the influence of the spatial component by showing one large stimulus covering all neurons’ RFs and requiring the animal to perform the task without considering the location of the stimulus (as opposed to using saccades as the behavioral response). Figure 2-11. A comparison with previous studies. (a) Different recording techniques leading to different spatial coverage and resolution. Our study used the distant paired recordings, which made sure that recorded signals represented the activity in different neural modules whereas the local paired recordings might pick up signals from the same module. The extracortical recording methods suffer from low spatial resolution and thus the 47 signal hardly reveals the feature-specific activity pattern. The neural module image was adapted from Roe et al., 2012 and used here for demonstration. (b) The traditional visual search task versus the feature match-nonmatch task used in our study. The visual search task confounds spatial and feature-related information, whereas our paradigm only focuses on the feature domain. The demonstration of visual search task was adapted from Bichot et al., 2005. Last, by analyzing the spike-field PPC, we were able to uncover the feature-dependent and attention-driven synchrony in different frequency bands. Many recent studies have focused on the stimulus-independent covariability between neurons, i.e., the noise correlation (K. Clark & Noudoost, 2018; Ponce-alvarez, Thiele, Albright, Stoner, & Deco, 2013; Ruff & Cohen, 2014). While they provide very interesting insights into the population code and/or the effective connectivity under top-down modulation, it has been argued that the spike count correlation mostly represents the low-frequency synchrony in the oscillatory network. In addition, the estimation of noise correlation requires higher number of repetitive trials, which was not possible in our experiment setting. Together, the choice of using spike-field synchrony as a measure of functional connectivity in our study enabled us to examine the synchrony phenomena in lower as well as higher frequency bands with a limited number of trials. 2.3.2 Common input or recurrent processing? There remains one critical question that our experiment did not answer: what is the origin of the identified synchrony pattern? Does it represent the in-region lateral interaction that is modulated by attention, or is it merely caused by a common driving input from another brain region, such as the frontal eye fields (FEF)? As we reviewed 48 earlier, there is evidence in support of both possibilities. In V4, neurons send axons to neighboring columns in a periodic manner, which implies that recurrent projections exist between neural modules (Yoshioka et al., 1992). On the other hand, studies have found bidirectional anatomical connections between V4 and FEF (Ungerleider et al., 2008). It was also reported that during spatial attention modulation, the gamma-coherence between FEF and V4 increased when the pair of recording sites represented the attended location, and this coherence was dominated by the top-down direction (FEF to V4). One could thus argue that the attention- related synchrony in V4 might just reflect the pattern of the top-down projection. Even if this is true, however, it is still not clear whether the local circuit plays a role in orchestrating the synchrony pattern of the distributed network in V4. Several statistical techniques were proposed to solve similar problems, such as considering the coherence part with non-zero phase lag (e.g., phase lag index, PLI) or autoregressive model-based analyses (e.g., Granger Causality). However, these methods do not rule out the possibility of top-down inputs with various time delays. Moreover, a heavily-interconnected recurrent network without a clear hierarchical structure might produce zero-phase-lag coherence by itself. Therefore, statistical analyses may not be able to disentangle these competing ideas. One reliable way of testing these hypotheses is to causally probe the functional connectivity using neural stimulation, such as optogenetics. Optogenetics has significant advantages compared with other stimulation techniques, including its fast responsiveness, cell-type specificity, and low electrical artifact which makes it 49 possible to stimulate and record simultaneously. Figure 2-12 demonstrates our previous attempt to carry out optogenetic tests using a multielectrode opto- microdrive, a modified version of GMR SC-32 with 4 optrodes (electrodes glued with optic fibers; see Figure 2-12a and Brooks & Sheinberg, 2015 for details) replacing the regular electrodes on the array. Through the same recording chamber in monkey X, we injected AAV virus carrying C1V1(an excitatory rhodopsin) with the CamKII promotor (targeting predominately excitatory pyramidal cells) at the optrode locations (as marked in Figure 2-12b). After 4 weeks of incubation time (virus expression time), we implanted the opto-microdrive and started to test the optogenetic effect using a 520 nm laser (with 2 mW power measured at the fiber tip) while the monkey was performing the delayed feature-match-nonmatch task. Figure 2-12(e) shows data from an example session. A 300-ms 50-Hz oscillatory optical stimulation was applied at channel 9 during the delay period. As channel 4 showed the largest stimulation effect instead of channel 9, we considered it as the “source” for our functional connectivity measure and calculated the coherence between channel 4 and all other channels on the array (shown as spectrograms in Figure 2-12e). It can be seen that some channels (even distant ones) show a bright spot at the stimulated time and frequency, which represents an LFP activity in synchronization with the stimulated site. Importantly, the phase lag for this coherence spot (marked with red arrows) is usually non-zero, indicating that the response is not merely caused by light spreading or volume conduction, but might reveal the underlying signal transmission through recurrent projections. Although this project did not yield enough data to draw any firm conclusions due to hardware 50 durability issues, we believe that, ideally, the same probing experiment could be repeated across different attention conditions in order to evaluate the attentional regulation of functional connectivity patterns. Figure 2-12. Simultaneous optogenetic stimulation and recordings. (a) The tips of an optrode under microscope without laser (upper picture) and with laser onset (lower picture). The laser transmission was tested with the 51 optrode inserted in a piece of agar gel to simulate the environment in the cortical tissue. (b) The array layout of the opto-microdrive on a brain model. The magenta circle represents the chamber location, black dots represent the electrodes, and white dots represent the optrode sites. The yellow shading on the brain marks the V4 region. (c) Injection of virus into the planned optrode locations through a grid. (d) The opto-microdrive implanted in the chamber after the expression of C1V1 (4 weeks). (e) An example session with coherence analysis. The stimulation was applied to channel 9 whereas channel 4 showed largest stimulation effect. Therefore, we show the coherence between channel 4 and other channels when a 50 Hz stimulation was delivered during the delay period of the task. The bright spots around 50 Hz / 0.5 s in some of the spectrograms represent the remote optogenetic effect. Red arrows denote the phase lag between channels, with the rightward representing zero phase lag. 2.3.3 What do different frequency bands represent? Our results revealed distinct synchrony patterns across different frequency bands, such as an inverted modulation between gamma and low-frequency bands, which is consistent with a large body of studies. One might ask, what (if anything) does each of these components signal? Presumably emerging from the interaction between local inhibition and excitation (Wang & Buzsaki, 2012), gamma rhythmic activity is thought to serve as a gain effector for the sensory processing (Tiesinga, Fellous, Salinas, Jose, & Sejnowski, 2005). It has long been known that sensory inputs induce locally synchronous gamma 52 oscillations within the neural population that processes the stimuli, whose amplitude and timing can predict the perceptual and behavioral outcome (Gray, Konig, Engel, & Singer, 1989; Hoogenboom, Schoffelen, Oostenveld, & Fries, 2010). Moreover, top- down signals, such as attention, also drive the gamma synchrony in sensory areas at the same time with firing rate gain modulation (Bichot et al., 2005; Fries et al., 2001). The role of gamma in sensory processing was further confirmed in a recent study, which showed that optogenetic stimulation-induced gamma activity causally enhanced the sensory response as well as the detection performance on weak stimuli (Siegle, Pritchett, & Moore, 2014). Altogether, these studies all provide evidence for a sensory facilitation role of gamma activity. In addition, the gamma rhythm plays a role in coordinating the fine temporal relationship between neural populations. In the neuronal communication through synchrony theory, temporally coordinated bursts of spikes can facilitate or suppress interneuronal communication, depending on the phase relationship between the upstream and downstream neurons, and gamma activities provide a perfect carrier for this modulation (Fries, 2005). Consistent with this idea, synchrony is thought to offer a flexible mechanism of forming dynamic sub-networks, called functional ensembles, out of high-dimensional neural systems such as frontal areas of the neocortex (Buschman, Denovellis, Diogo, Bullock, & Miller, 2012). Despite the lower dimensionality in earlier sensory areas, gamma synchrony may still serve to bridge physically separate neurons within the same computational domain. As an example, an early study in cat V1 found zero-phase-difference gamma synchronization between spatially separate columns that share similar orientation selectivity (Gray et 53 al., 1989). The phasic coordination of gamma rhythm not only exists within one region, but also takes part in the communication between areas. Attention was shown to enhance the V1-V4 and FEF-V4 synchrony for specific neural pairs that correspond to the attended representation, which might facilitate the interareal communication efficacy (Gregoriou et al., 2009; Grothe et al., 2011). Taken together, gamma oscillation and synchrony play a critical role in the modulation of sensory processing. The low-frequency activities such as alpha and theta rhythms, on the other hand, are proposed to apply the top-down gating and/or sampling “guidance” to the sensory regions. Recent studies have shown that, while the gamma-band activity subserves the bottom-up information flow, alpha waves mainly represent the inhibitory top- down effect (Pelt et al., 2016; van Kerkoerle et al., 2014). Multiple studies have also shown a selective alpha desynchronization related to the cognitive engagement or increased alpha synchrony for the distractor representation (Bollimunta et al., 2011; Buffalo et al., 2011; Fries, Womelsdorf, Oostenveld, & Desimone, 2008). Importantly, such alpha modulation was found to be driven by feedback control from FEF, which precedes and predicts the amplitude of stimulus-induced gamma activity in the visual cortex (Popov, Kastner, & Jensen, 2017). Similarly, a decorrelation of the intrinsic theta fluctuation among neural populations was found during spatial attention modulation, suggesting an active upregulation of the signal-to-noise ratio (Mitchell, Sundberg, & Reynolds, 2009). In addition, theta and alpha oscillations were also shown to regulate the phase of “gamma envelopes” under various cognitive conditions, indicating a role of the cross-frequency interaction in the feedback control (Canolty & Knight, 2010). Despite controversies on the origin of these low-frequency 54 waves, it is commonly believed that they migrate over a larger cortical range compared with gamma activities, represent the internal cognitive state, and interact with the sensory processing through a top-down flow. Returning to our study, we showed a low-frequency synchrony among spatially- distant neurons with similar feature-tuning in V4 in the absence of visual stimuli, whereas gamma synchrony emerged within these feature-dependent subnetworks when stimuli were presented. Based on our previous discussion, this pattern seemed to demonstrate how the bottom-up input may disrupt the default low-frequency- dominant pattern with synchronous gamma activities. Moreover, our results further showed that feature-based attention increased the gamma synchrony between neurons that represent the attended feature, while reducing the low-frequency (theta) synchrony as well as the alpha-gamma phase-amplitude coupling. Consistent with the role of gamma synchrony in spatial attention modulation, our findings suggested that feature-based attention facilitated the interaction within the task- relevant neural ensemble through gamma synchrony, which might contribute to the gain modulation as well as an increased efficacy of the bottom-up processing. Meanwhile, it is worth noting that the gamma modulation occurred later than those effects in the low-frequency range, which support the idea that low-frequency desynchronization leads the top-down modulation. 55 2.3.4 Potential roles of the tuning- and attention- dependent gamma synchrony We mentioned above that the gamma synchrony might serve to selectively modulate the local and long-range communication efficacy. How does this mechanism contribute to the information coding in a network? What is the computational role of the tuning- and attention-dependent synchrony found in previous and current studies? One idea is that the gamma synchrony pattern forms a flexible recurrent network that enables the attractor network property in both spatial and feature dimensions. When attention is driven toward a certain feature or location, the change of the synchrony (effective connection pattern) transiently improves the representation of the relevant information. To test out our idea, we propose to explore the network dynamics and information coding using a simplified recurrent neural network model that has an implementation of neural oscillation and synchrony. First raised by Reichert & Serre, 2013, the complex-valued Boltzmann machine models may provide an ideal starting point for our exploration. In Boltzmann machine models, units form two or more computational layers with symmetric bi-directional connections. The first layer represents the pixel information of the input images (the “pixel” layer), and the second and higher layers serve as hidden layers, which are trained to represent the learned images in the lowest energy through recurrent processing, so as to uncover the statistics of images. 56 Reichert & Serre, 2013 replaced the network nodes with complex-valued units 𝑟𝑟𝑖𝑖 𝑒𝑒 𝜓𝜓𝑖𝑖 , in which the magnitude 𝑟𝑟𝑖𝑖 stands for the average firing rate of the unit, and the angle 𝜓𝜓𝑖𝑖 corresponds to the current phase of the oscillation. Such a design allowed the phase relationship (synchrony) to be part of the activity evolution and information coding. They found that after proper training as regular (real-valued) networks, the complex-valued Boltzmann machine not only represented the features of the training images in a similar way to the real-valued ones, but also naturally showed the interesting binding-by-synchrony phenomena. For example, when multiple stimuli were presented to the network simultaneously, the pixel layer units that encode the objects with similar features tent to be phase-synchronized (as shown in the left column of Figure 2-13), presumably due to their bidirectional interaction with the higher-layer feature-selective units. This model showed how feature-based synchrony can be explained using a simple neural network model and a simple learning rule. We propose to extend the complex-valued Boltzmann machine with attention mechanisms so that it can be used to explore the computational role of the synchrony modulation underlying attention. Specifically, we could train a 3-layer network with a pixel layer, a hidden layer, and a semantic layer that represent the feature/spatial labels using regression or supervised learning methods (Figure 2-13). Attention will be implemented with an increased activation amplitude of a feature-specific or spatial-specific unit in the semantic layer (i.e., the red circled units in Figure 2-13), corresponding to the feature-based and spatial attention, respectively. We expect that: (1) the “injection” of attentional signal will lead to enhanced synchrony between 57 middle-layer and pixel-layer units that encode the attended objects (those sharing the attended feature or located at the attended location) (2) the modulation of synchrony pattern will improve the representation of the attended information (e.g., increased decoding performance or robustness to noise). We hope that this proposed work will provide a unified model for feature-based and spatial attention, which allows us to experiment in a well-controlled setting and explore the potential computational roles of attention-related synchrony in the network. Figure 2-13. A conceptual model for exploring the role of synchrony in attention modulation. The top row shows the structure of a 3-layer Boltzmann machine model, in which lines represent connection weights, and circles representing units with their represented features (e.g., L for left, R for right, bars for different orientation). The middle row shows the input pattern (bars with two different orientations and randomized locations). The bottom row shows the hypothesized outcome at the pixel layer after recurrent processing, in which the color represents the phase of the corresponding “neurons”. The 3 columns represent 3 conditions, i.e., without attention, with feature-based attention (activation of a feature-coding node in the 58 semantic layer), and with spatial attention (activation of a spatial-coding node in the semantic layer). 2.3.5 The effect of working memory Attention modulation usually shares a tight link with working memory, i.e., the temporary storage and manipulation of information related to the current task. It is commonly believed that working memory is maintained through sustained spiking activities, or more generally, through delay activities (Fuster & Alexander, 1971; Lundqvist, Herman, & Miller, 2018). Along the visual hierarchy, it has been found that such delay activities in relation to feature-based information emerge sharply in the late stages (e.g., MST, IT, and LPFC) but not as much in early and mid-level visual areas (e.g., V1, V4, and MT), suggesting the role of association regions in the working memory retainment (Leavitt, Mendoza-halliday, & Martinez-trujillo, 2017; Mendoza- halliday, Torres, & Martinez-trujillo, 2014; Woloszyn & Sheinberg, 2009). However, there remain debate on whether working memory processes require a distributed network involving earlier sensory regions, given that some studies have found modulations in spiking activity and/or LFPs during the delay period in these regions, despite a much smaller amplitude than in association areas (S. J. Luck et al., 1997; Mendoza-halliday et al., 2014). One possibility is that during the delay period, top- down signals that carry the memorized information project to earlier visual stages and affect the sub-threshold state of neurons, which can be shown as a small modulation on spiking and LFP activities. These modulations, instead of coding the 59 memory information explicitly, might play a role in optimizing the processing system according to the memory content for further sensory tasks (Merrikhi et al., 2017). The delay period in our paradigm could have provided a chance for us to investigate the effect of working memory related to the attended feature. However, we did not observe robust sustained activities or a significant modulation on the synchrony pattern during the delay period. While there are a number of possible explanations for this lack of delay activity, we think two factors may be critical: the signal-to-noise ratio and the length of the memory period. Since it is commonly reported that the working memory related modulation can be as small as 1 Hz (in terms of firing rate), we expect that a large dataset, with extremely high signal quality and optimal task conditions, may be necessary to capture the small effect size. These conditions might not be fully satisfied in our dataset. In addition, studies of working memory generally use a longer delay period (e.g., 1 second or above) to separate the memory period from the sensory processing and to ensure the reliability of spectral analyses in low- frequency bands. Our paradigm only includes a 500-ms delay period, with the major purpose of separating the visual probe (attention) period from the visual processing of the sample. Due to the limited time length for each session (how long the animal can perform the task per day) and our demand of a large trial repetition number (how many trials of experiment we need to accomplish per session), our task design necessarily reduced the overall trial length, including the length of the delay period. Future studies could, of course, use the same feature-match task with a longer delay period in order to investigate the role of working memory-related feedback in early/mid-level visual areas. 60 2.3.6 Obstacles and lessons Throughout the project, we came across many challenges and learned lessons through mistakes. We would like to keep a record for other people to avoid detours in their future projects. (1) Task design. In our early attempt for this project, we adopted the traditional cued visual search paradigm with an array of 2 small stimuli (colored texture patches). The monkey was asked to make a saccade to the only stimulus that shares the same feature with the cue (Figure 2-14a). This design is ideal for single-electrode or small-scale recordings, where the receptive field and feature tuning of the recorded neuron is measured carefully and used for configuring the task stimuli. When it came to multi-channel recordings over tens of neurons, however, we failed to tune the task parameters to the preference of all (or most) recorded neurons, making the experiment less efficient. In order to cover the neurons with diverse preferences, we had to increase the number of unique conditions, and that reduced the number of repetitions to the point where pairwise synchrony analysis became unreliable. What is worse, some stimuli, while showing up near the center of some neurons’ receptive field, appeared around the border of other neurons’ receptive field (as demonstrated in Figure 2-14a), which led to unwanted spatial/border-related response. Moreover, the task induced a transformation from feature-based attention to spatial attention. Using a support vector machine (SVM), we could decode the information carried by population activities (Figure 2-14b), and we found that the late visual 61 period (before saccade) was dominated by the spatial information of the target, leaving the feature-based attention effect barely decodable. In addition, the delay period (and perhaps the search period as well) in our first version of task was largely contaminated by the activity in the cueing period. The cue was initially designed to appear in the foveal location, which was close enough to activate some recorded neurons. The delay between the cue and search array (300 ms in the first task) was not long enough to separate the two visual periods. We could not rule out the possibility that the attended feature information decoded from the population activity was merely an aftereffect from the visual response to the cue. 62 Figure 2-14. Our original design of task. (a) The visual search task. The monkey was required to keep fixation at the central dot while a cue (color or texture) was presented at the foveal location and following a delay period, an array of 2 colored texture patches showed up, within which only one stimulus (the target) shared the same feature with the cue. The monkey should wait until the visual stimuli disappeared and make a saccade to the location of the target. White dashed circles demonstrate the common receptive field locations of the recorded neurons. (b) Example results of the population neural decoding using spikes and LFPs for the cued feature information (left) and the target location information (right). Each trace represents the decoding accuracy using a sliding window over time, and the error bar represents the cross-validation variance. 63 For all these reasons, we switched to our current task, the cued feature-match- nonmatch task, with the cue appearing opposite to the neurons’ preferred visual space and the large visual probe covering all their receptive fields. Such a design minimized the effect of spatial component and reduced the number of conditions required for probing the recorded population. We also slightly increased the duration of the delay period as well as the visual period, making the synchrony estimate more reliable (as initial visual peak period did not satisfy the stationary assumption) and less affected by the cueing visual response. The exploration of a better paradigm taught me that “less is more”, i.e., a simpler design with fewer components may bring more power to discoveries and statistics. (2) Chamber implantation. The chamber implantation/ maintenance is never a trivial task in a monkey lab. In our case, we used the Gray Matter Research tangential chamber, and challenges mainly came from two aspects: a) the location of V4, b) the use of Metabond®. The V4 area, especially the near-foveal region (our target), is located close to the monkey’s ears (Figure 2-15). For one thing, it is critical to remove a significant amount of temporal muscle when making the craniotomy to insert the recording chamber, as shown in Figure 2-15(b). We failed in our first implantation attempt because we were conservative about the muscle removal, and as a result, the swelling muscle climbed over and buried the newly implanted chamber after surgery. Also, the large angle of the chamber makes it difficult to apply cement material (e.g., Metabond in our 64 experiment) on the ventral side of the chamber. It caused sealing failure in one of our implants, which led to a long-lasting infection on the dura even if we fixed the sealing with more acrylic in a follow-up surgery. It is also worth noting that the security of Metabond highly depends on the dryness of the bone surface as well as the temperature of the Metabond. Taken together, we learned that we should not underestimate the importance of a clean and secure surgical target, especially when placing a self-contained and semi-chronic device that cannot be easily cleaned (such as an open chamber). If possible, make larger skin opening, remove the tissue and muscle that might interfere with the craniotomy, clear the surrounding surface with hydrogen peroxide and gauze, and layer the Metabond carefully around the chamber bottom to provide a seamless sealing. Figure 2-15. A demonstration of the V4 chamber localization. (a) The chamber and electrodes location on the inflated brain model reconstructed from the MRI scanning result. (b) The center location of the chamber on a coronal view of the monkey brain. 65 (3) Recording setup. We selected to use the GMR semi-chronic microdrive system for recordings due to its balance between durability, flexibility and ease of use, and also for its large spatial coverage. However, our experience was not ideal and provided lessons for future experiments. First, the penetration of electrodes into the brain requires dura to be fresh and thin. We broke more than half of the electrodes in our first implanted drive, because we waited over 1 month before we began lowering electrodes and the dura was significantly thickened. Second, the inner space of the drive needs to be watertight sealed. We had several drives that started to show electrical shortcut or bad contact artifacts within one month due to liquid damage on the printed circuit board. We thus sealed the cap with Teflon tape and/or silicone gel to avoid leakage. Third, the electrodes are not always safe even when they are in the cortex. The dura and the brain tissue fluctuate vertically to a surprisingly large extent, especially during the post-operation days or when the animal is back on water restriction. We eventually had to lower most electrodes deep enough such that they would retract beyond the dura with the brain fluctuation. Fourth, although the electrodes are designed to be adjustable for obtaining well-isolated single unit signals each day, the range for exploration (i.e., the depth of the cortex) is not big and we could run out of travel distance easily. Retracting the electrodes and revisiting the depths that were explored before may work in some cases, but that usually reduced the chance of finding well isolated cells. Therefore, for most of our recordings, we kept the electrodes nearly stationary unless they no longer picked up spikes. Fortunately, our daily analysis on the neurons’ tuning properties indicated that the same electrode 66 usually sampled different neurons on different days even if its location was not adjusted by the experimenter. (4) Synchrony measure. As to the synchrony evaluation, before turning to the spike-field PPC, we were initially focused on the LFP-LFP coherence and used the LFP response pattern to characterize the feature tuning. Despite our success in analyzing receptive fields with the gamma- band LFP power (described in section 2.2.2), the same method did not perform as well in the feature domain. LFP channels either showed weak modulation by features or showed highly correlated patterns between channels. In addition, we did not find significant modulation of the LFP-LFP coherence based either on the LFP tuning or the spike tuning. Therefore, we concluded that the LFP signals might not be an ideal readout for the population coding of feature, and instead, we decided to estimate the synchrony using the spike-field PPC with spike-defined tuning properties. It turned out that these measures worked well, even though the number of available channels was limited by the quality of the spiking data. (5) Large data handling. The multi-electrode recording produced GBs of data each day, which initially overwhelmed our platform and our skills for dealing with the data. For example, loading and preprocessing even a single session would take minutes of time. Working together with our colleague Shaobo Guan, we developed a Python-based platform for handling, analyzing and visualizing electrophysiological data. We reorganized the original continuous data from each session into arrays of epochs, appending the 67 channel and task information, stack multiple sessions, and stored the whole structure in HDF5 format. HDF5 is ideal for organizing big data, not only because it is compatible with both numerical and string data, but also because it can be loaded in segments. For instance, we could load the data from several specific days, or only read the data tagged with “spike” but not “LFP”. This allowed high flexibility and efficiency when exploring the data. Beyond data storage, we also took advantage of the powerful computation server provided by the lab and tried to parallelize any expensive computation, such as PPC estimate. All these practices made it possible to dig into the large data sets we collected. (6) Effect size. Overall, the effect size for our findings was not large, even the attentional effect on single-neuron level. One possibility may be that the stimuli used in our experiment were not optimized for all neurons’ tuning preference (due to the tradeoff for a simpler experimental design), yet the strongest attention modulation is usually applied to the neurons that prefer the attended information (e.g., Bichot et al., 2005). Another reason might be related to the signal-to-noise ratio for the PPC estimate. On one hand, the relatively low firing rate in our data (perhaps due to the use of suboptimal stimuli) raised the estimate variance of spike-field PPC. On the other hand, although we made our best efforts to increase the sample size by reducing the condition number, it might still call for a larger number of repetitive trials to ensure a more reliable spectral phase analysis. (7) Problems with one animal. 68 The most significant problem for this project is that we were not able to show the same effect in each of our two animals. The amount and quality of the data from monkey X were both affected by our attempts at using optogenetic manipulation to probe the functional connectivity of area V4 before collecting the data presented in this thesis. We collected data from monkey X at the same time that we performed the very first test for the GMR opto-microdrive system. Our plan was to carry out the optogenetic probing for the functional connectivity while investigating the non- stimulation activity pattern, all with the same setup. However, the test of a newly designed device as well as the development of a new protocol were unexpectedly challenging. One challenge came from the injection of virus in a large area of cortex since the drive has 4 distributed stimulation sites. We kept the chamber open, repeated viral injection at multiple locations, which took more than 1 week, waited for ~4 weeks, and tested the expression using acute single electrode recordings until we implanted the chronic drive. Another challenge was the penetration of the two- tipped optrodes through the dura. We performed micro durotomy during the surgery to leave passages for the optrodes and drove them down immediately after the surgery in the operation room. All these manipulations, however, seemed to interfere with the health of the dura and the cortex. We observed a dramatic change of the dura depth and its thickness once we closed the chamber with the chronic drive. The suspected immune activity around the dura pushed the cortex down, which caused a shortage of travelling length for our recording electrodes. In the end, we had collected only 11 sessions from monkey X (compared with 23 sessions from monkey H), and the number of neurons recorded in each session was also smaller, leading to a 69 substantial limitation on the available data (e.g., only 7 eligible signal pairs for the attention effect analysis whereas the number from monkey H being 68). It is therefore unfortunate that we could clearly confirm our findings in more than one animal, but rather show the great potentials with mainly one animal’s data. 2.4 Method 2.4.1 Animal manipulation Two adult male macaque monkeys, X (10.0 kg) and H (9.6 kg), were used in the experiment. Before implanting any metal on the monkeys, we first performed an MRI scan to establish a 3-D structural model of their brains. Magnetic responsive labels were attached to the plastic implants on their skull during the scanning to serve as landmarks for future calibration. After training on behavioral tasks, each monkey was implanted with a titanium chamber and a semi-chronic SC-32 microdrive (Gray Matter Research) above their V4 areas (left hemisphere for monkey H and right hemisphere for monkey X). Chamber localization was guided by the aforementioned MRI scanning result using the Brainsight calibration system. During training and recording days when the animals were water restricted, their weight and other health signs were closely monitored on a daily basis. All procedures approved by the Brown University IACUC and following NIH guidelines for animal care and use. 70 2.4.2 Behavioral paradigm The behavioral paradigm is illustrated in Figure 2-2. In the feature-match task, a trial starts with a feature sample (4° in diameter) appearing at a near-foveal location (3° to the left, and 3° to the top). The sample has two possible forms, either a gray-scale grating patch or a colored Gaussian blob, which indicated the type of feature that the monkey should keep in mind on that trial. The initial sample stimulus remained visible for 300 ms, and was followed by a 500 ms blank-screen delay period. After the delay period, a colored grating (the probe stimulus) was shown in the opposite quadrant of the screen. The probe was as large as 10° in diameter, with its center located at (4°, 4°) to the lower right direction, so that it covered most of the recorded units’ receptive fields. The monkey was trained to compare the visual probe with the remembered sample and determine whether the probe had the same orientation (if the previously seen sample was a grating) as the sample or whether it was of the same color (if the sample was a colored blob). The probe stimulus remained for 500 ms, after which time the monkey could report the decision by pressing the corresponding button (right button for “match” trials, and left button for “nonmatch” trials). A drop of juice was delivered on correct. To sum up, there were a total of 16 conditions: 2 feature types (orientation/color) * 2 match conditions (match/nonmatch) * 4 possible sample orientation * 4 possible sample colors. All conditions were counterbalanced, repeated ~125 times, and mixed into multiple blocks. The animal was required to keep fixation at a central point on the screen throughout each trial, and eye position was monitored using an Eyelink eye tracker with 500 Hz sampling rate. 71 In the spot flashing paradigm for identifying neurons’ receptive field, the animal was required to maintain fixation to a central point while a small white spot (0.5°) rapidly flashed at different locations randomly sampled from a large visual space. Each spot was shown for 50 ms, and 20 spots were presented each trial. To keep the animal alert, the fixation point jumped to a peripheral location at the end of a trial, and the animal was rewarded for making a saccade to the displaced fixation point. 2.4.3 Visual stimuli The visual stimuli shown in the task were generated using the STIM (a stimulus generating system) running on Windows 10, and the experimental task was controlled with the ESS state system running on QNX4 real-time operating system, both developed by Dr. David Sheinberg and colleagues. All stimuli were presented on a Display++ LCD monitor (68 cm * 38 cm) at a 100 Hz refresh rate, positioned in front of the animal with a 104 cm viewing distance. The color variants used in the task were sampled from the CIEXyY color space, with a fixed luminance (0.6 times the luminance of white) and saturation (0.6). The white color on the monitor was measured to be 108.1 cd/m2, and the gray background was 54.3 cd/m2. 2.4.4 Data acquisition and processing All recordings were conducted using the Gray Matter Research (GMR) semi-chronic SC-32 microdrive loaded with 32 Alpha Omega glass-coated tungsten electrodes (125μm, ~1MΩ). In common cases, we were able to record from 10~18 working channels simultaneously, and the depth of the electrodes could be adjusted daily to obtain new units. The signal was amplified and digitized by the TDT digitizer PZ2 at 72 50 kHz sampling rate and processed using the TDT data acquisition system RZ2, through which the raw signal was band-passed into a low frequency part (1~300 Hz) and a high frequency part (>3000 Hz). LFPs were down-sampled at a 1017 Hz rate from the low frequency component of the original signal. Spikes were identified from the high-frequency component using a threshold of 3.5 times of the standard deviation of the signal. We sorted the spikes from different units using the Plexon offline sorter. However, due to the relatively low signal-to-noise ratio, we considered both the MUA and SUA as spiking units in our analysis. The recording data for this project came from 34 sessions (11 from monkey X, and 23 from monkey H), where 23 were full sessions (9 for monkey X, 14 for monkey H) and the other 11 (2 for X and 9 for H) ran orientation-match task only (i.e., no color trials). There were in total 636 spiking units identified and put in the dataset, plus the LFP signals from the same channels where these spiking units came from. The spike time stamps were binned so that it shares the same length with the LFP data (using the same sampling rate). The binned spike rates were then smoothed using a Gaussian kernel (sigma=10 ms) and normalized based on the mean visual response rate (50- 250 ms post stimulus onset). Both spike trains and LFPs were aligned to the sample onset. Data were arranged and stored for further analysis using the Python-based analysis toolbox PyNeuroSG developed in collaboration with Shaobo Guan (https://github.com/SummitKwan/PyNeuroSG). 73 2.4.5 Feature tuning and receptive field measurement Before analyzing single neuron activities, we first judged whether a neuron was visual or non-visual by comparing its activity during baseline (-200-0 ms before trial onset) and the visual peak (50-250 ms post stimulus onset) periods in the delayed feature match task. The spike count in both periods were calculated for each trial, and their between-group difference was measured using Cohen’s d method: 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀(𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣) − 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀(𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏) 𝑑𝑑 = ��𝑉𝑉𝑉𝑉𝑉𝑉(𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣𝑣) + 𝑉𝑉𝑉𝑉𝑉𝑉(𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏)�/2 A neuron was denoted as a “visual cell” if d>0.5, and non-visual cells were removed from the pool in further analyses. For all visual cells, their tuning properties were identified using the response rate to different visual probes in the delayed feature match task. Spikes within the 50-400 ms time window post stimulus onset were counted and placed in a 2-D matrix that represents the orientation and color feature dimensions of the visual stimuli. The selectivity domain type was defined as the dimension along which the matrix showed a more significant modulation. Specifically, we performed one-way ANOVA along each dimension and divided the neurons based on the F and p values, using different criteria for domain-level analyses and specific feature-level analyses (Table 2) in order to increase the data amount for the second set of analyses. Once we identified the domain type, we then assigned the preferred orientation for each orientation- selective neuron and the preferred color for color-selective neurons, based on the peak of the 1-D tuning curves. 74 Preferred-feature-dependent Domain-based synchrony analysis synchrony analysis & Attentional effect analyses 𝑃𝑃𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 < 0.001 & Both type 𝑃𝑃𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 < 0.001 & (sensitive to both |𝐹𝐹𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 − 𝐹𝐹𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 | N.A. features) < 0.5 𝐹𝐹𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 + 𝐹𝐹𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 Orientation- 𝑃𝑃𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 < 0.001 & does not 𝑃𝑃𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 < 0.001 selective type belong to the “both” type Color-selective 𝑃𝑃𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 < 0.001 & does not belong 𝑃𝑃𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 < 0.001 type to the “both” type Does not belong to any type None type Does not belong to any type above above Table 2. The criteria of categorizing neuron tuning types for two sets of analyses. We also estimated the receptive field (RF) of neurons and LFPs using the spiking and gamma-power (35-55 Hz) in response to visual probes at different locations in the dot flashing viewing task. The time window for calculating the response was 50-250 ms. The response distribution for each signal was then smoothed with a 2-D Gaussian Kernel (sigma=1 ̊) and filtered with a threshold of half amplitude (peak-to-peak), resulting in a 2-D matrix of RF. The center of each RF was identified as the center-of- mass of the RF matrix. 2.4.6 Pairwise synchrony analysis The pairwise synchrony was measured as the spike-LFP pairwise phase consistency (PPC) (Vinck et al., 2010). Proposed as a bias-free metric, PPC has been widely used to evaluate the synchrony between neuronal spikes and LFP without suffering from the influence of different spike counts and trial numbers. It statistically analyzes the phase of the spike timing relative to the LFP oscillations (in some frequency band), 75 similar to the idea of phase locking value (PLV) or coherence, but instead of taking the population vector average, it considers the statistics of pairs of spikes using the average pairwise circular distance (APCD): 1 �= 𝐷𝐷 � 𝑑𝑑(𝜃𝜃𝑖𝑖 , 𝜃𝜃𝑗𝑗 ) 𝑁𝑁𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 𝑖𝑖,𝑗𝑗 where 𝑑𝑑(𝜃𝜃𝑖𝑖 , 𝜃𝜃𝑗𝑗 ) is the absolute angular distance between the relative phase of each pair of spikes. The APCD is then normalized to fit in the range (0, 1). It has been shown that the PPC is equivalent to the statistics of squared PLV, and importantly, it breaks the finite sample size bias, i.e., the resultant value of phase relationship tends to be smaller with larger sample size, as in PLV or coherence estimates. To further reduce the uneven sampling problem and increase the robustness, we selected to use a variant of PPC estimate, which excludes spike pairs coming from the same trial. The PPC analysis was carried out using the MATLAB toolbox “FieldTrip” (Oostenveld et al., 2011), including the estimation of the spike-triggered spectrogram using a frequency-specific Hanning window of (2/frequency+0.15) s and the PPC statistics with window size equal to 0.2 s. The frequency-specific time window was selected to resolve the tradeoff between temporal resolution and estimation robustness. For each condition of interest, we also calculated the PPC spectrogram using the shuffled data (trial index randomly permutated for the LFP data while keeping the original trial assignment for the spike data) to remove the trial-evoked component of synchrony. 76 The cross-frequency phase-amplitude coupling (PAC) was performed using the python-based toolbox “pacpy” developed by the Voytek Laboratory (http://voyteklab.com), which calculates the coherence between a lower-frequency signal and the envelope of a higher-frequency signal. In each trial, the raw LFP from two channels were band-passed using finite impulse response (FIR) filters. The lower-frequency signal (lo) was filtered in the theta band (3~5 Hz) or the alpha band (8~10 Hz) with w = 2 cycles, and the higher-frequency signal (hi) used the gamma band (35~60 Hz) with w = 3 cycles. The difference between the phase of the lower- frequency signal and the amplitude of the higher-frequency signal was given by: 𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 = 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎[𝐻𝐻(𝑙𝑙𝑙𝑙)] − 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎[𝐻𝐻(|𝐻𝐻(ℎ𝑖𝑖)|)] where H(∙) stands for the Hilbert transformation, angle(∙) and |∙| represent the phase and the amplitude of the oscillatory data. The resulting diff was then used to calculate the inter-trial coherence: 𝑃𝑃𝑃𝑃𝑃𝑃 = � 𝑎𝑎𝑎𝑎𝑎𝑎(𝑒𝑒 𝑖𝑖∙𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 )� To exclude the effect of trial-locked activity, we then repeated the calculation of PAC with one out of the two signals shuffled across trials within each condition, and the permutation result was removed from the original PAC result. 77 Chapter 3 The neural and behavioral dynamics of feature-based attention 3.1 Motivation The thesis so far has focused on the oscillation synchrony pattern underlying the feature-based attention modulation in visual area V4, yet a major question was not directly addressed in this study: how does the neural oscillation and feature-based attention affect behavior? Thus far, our paradigm was designed primarily to maximize the efficiency for multi-electrode recordings, but at the same time, it was lack of a simple measurement of the behavioral output. There is a large body of literature using threshold-level detection tasks to evaluate the behavioral effect of neural oscillation and attention (e.g., a review by Buschman & Kastner, 2015). Through correlating the detection performance with the phase of internal oscillations, it has been suggested that spontaneous or attention-induced oscillations, especially in lower frequency bands, modulate the rate of successful detection. More interestingly, spatial attention is shown to improve the behavioral robustness by suppressing the behavioral phase dependency on oscillation (Harris, Dux, & Mattingley, 2018). Here we ask if feature-based attention produces a similar effect. 78 Figure 3-1. The attention-related behavioral oscillation. (a) In the study by Fiebelkorn, Saalmann, & Kastner, 2013, human subjects showed an oscillatory pattern of detection performance as a function of the cue- target interval, which might represent the neural oscillation underlying the attention-induced rhythmic sampling (figure adapted from Fiebelkorn, Saalmann, & Kastner, 2013). (b) In the study by Harris et al., 2018, detection performance depended on the phase of the ongoing theta- and alpha-band oscillations, and such phase dependency was suppressed when the target appeared at cued (attended) locations (figure adapted from Harris et al., 2018) . We carried out this project by combining human psychophysics and electroencephalogram (EEG) recordings, with the aim of building a link between the neural and behavioral oscillations under feature-based attentional modulation. Meanwhile, this large-scale noninvasive recording technique could also allow us to zoom out of one single brain region and explore the whole-brain level network. We tested 8 subjects in and we believe that these initial results can provide some insights to complement to our understanding of the dynamics of feature-based attention. 79 3.2 Results 3.2.1 Experiment design In this project, we adopted the threshold-level detection paradigm to probe the neural oscillatory activity (Figure 3-2a & b). The task was designed with the following considerations: (1) To make the detection performance more sensitive to the change of internal state, we reduced the amount of sensory information and kept the detection rate around 50%~60% by introducing a noisy presentation pattern. Sitting on top of a flickering dot noise background, the detection target was shaped by a local change of the color statistics of dots, and thus the level of difficulty could be controlled using the color coherency parameter. (2) To ensure that subjects were actively detecting, we used a relatively high rate of catch trials (20%) where the target was absent. The performance for these trials showed a reasonably low false alarm rate in Figure 3-2(d). (3) To engage feature-based attention, we presented a cue before the target onset to indicate the possible color of the target with 2/3 validity (the “attended” condition), whereas in 1/3 of non-catch trials (the “unattended” condition) the target appeared in the non-cued color. (4) To reduce the potential expectation of timing, and also to evenly sample the different neural phases, we varied the cue-target time delay within a 250 ms range (400 – 650 ms). 80 (5) To make sure that the attentional effect on the behavioral oscillatory pattern was not merely due to different levels of performance (such as the ceiling effect potentially reducing the behavioral oscillation in the attended condition), we adjusted the detection difficulty for the attended condition and unattended condition independently so that the performance was balanced. Figure 3-2(c) shows the difficulty value (i.e., 1 - color coherency) used for each subject. The difference between attended and unattended conditions represents the benefit of feature-based attention. As shown in Figure 3-2(d), after proper training and manual adjustment of difficulty values, subjects could perform the task with a 50%~60% detection hit rate for both attended and unattended conditions, with false alarm rates of approximately 20%. Figure 3-2. The task design and behavioral outcome. 81 (a) The time course of the cued detection task (please refer to 3.4.2 for a detailed description). (b) A demonstration of different trial types and their proportion. (c) The difficulty value (i.e., 1 - color coherency) used for each subject. Each gray line connects the difficulty in both conditions for the same subject. (d) The subjects’ performance for attended, unattended and catch trial conditions. 3.2.2 The spatial-temporal structure of EEG response We recorded the EEG signal while subjects performed the cued detection task. The recording electrodes mostly covered occipital and parietal regions. Figure 3-3 shows the time-frequency power spectrogram of the post-cue activity averaged across channels and subjects, with the mean baseline (-0.2 – 0s) power deducted. The cue induces a transient increase of theta-band power which might represent the event- related potentials (ERPs) in response to cue presentation. In addition, there is a long- lasting decrease of alpha-band oscillation which might be related to the cognitive state triggered by the cue onset. Figure 3-3 also shows the topographic distribution of the theta and alpha powers in several time slices. It can be seen that the alpha power decrease is most prominent in the occipital and occipital-temporal channels, whereas the visual-induced theta power is found in the occipital area first and then appears concurrently in a central frontal-parietal area. This result provides the first confirmation that our task modulates the EEG signal in a global manner. 82 Figure 3-3. The topographic distribution of post-cue oscillation power changed with time. The figure superimposes a time-frequency power spectrogram averaged among all channels and the topographic distribution of the power in alpha and theta frequency bands at multiple times. The color in both cases represents the log- ratio of the power at each time/frequency relative to the baseline (-0.2 – 0s) power of the same frequency. 3.2.3 Detection performance depends on the peri-stimulus alpha phase To explore the relationship between behavior and neural oscillations, we adopted the phase dependency testing method used in Harris et al., 2018. Basically, it assumes that if the oscillation phase modulates the behavioral performance, we should expect to see different clustering in the phase distribution for one condition (e.g., hit trials) compared to another (e.g., miss trials), and such clustering should significantly differ from the results after a shuffled permutation. For each subject, we took a pair of focal 83 electrodes corresponding to those in Harris et al., 2018, and computed the phase opposition summation (POS), a measure for the phase clustering. Figure 3-4 shows the P value for the permutation test on POS as a function of time and frequency, for the attended and unattended trials, respectively. A bright spot means that the phase distribution at that certain time and frequency significantly differs between hit and miss trials. It is found that under both attended and unattended conditions, the alpha- band phase right after the target onset seems to be correlated with the detection performance, i.e., there is a higher chance for successfully detecting the weak target when the target appears at a “good” phase of the alpha oscillation. Based on this result, we will focus on the time-frequency window where both conditions show a significant modulation (frequency = 8 – 14 Hz, time = 0 – 0.1 s). 84 Figure 3-4. Log(p) values of phase dependency as a function of time and frequency. (a) Attended condition. (b) Unattended condition. The colored heat map represents the log(p) value of the POS. Yellow dashed line marks the target onset time. 3.2.4 The phase dependency is suppressed by feature- based attention Having identified the time and frequency window in which the phase of oscillation is most influential, we then directly compared the phase dependency of performance between attended and unattended conditions. To show the phase dependency, we took all the phase data within the identified window (8 – 14 Hz frequency, 0 – 0.1 s post target onset), binned the phase values into 7 even bins, and calculated the 85 subjects’ detection rate for each phase bin. Figure 3-5 shows the average performance-phase function with the phase that has the highest hit rate (the “best” phase) aligned in the center. As we can see, the unattended trials showed a higher phase dependency than attended trials, as the performance difference between the “best” phase and “worse” phases is larger in the unattended condition. Importantly, we counterbalanced the number of trials in each plot, so that the different pattern was not caused by the larger data size for the attended condition. Our preliminary results seemed to indicate that, similar to spatial attention, feature-based attention also acts to suppress the neural rhythmic effect on the detection behavior. Figure 3-5. The alpha-phase dependency of hit rate under attended and unattended conditions. 86 Detection rate was averaged within each phase bin (8 – 14 Hz frequency, 0 – 0.1 s post target onset). The phase with the highest performance was aligned in the center. The red trace represents the attended condition, whereas the blue trace represents the unattended condition. Error bar stands for the inter-subject standard error. 3.2.5 Exploring the functional connectivity The large-scale recording using EEG also provided a tool for us to uncover the global interaction pattern evolving during the feature-based detection task. We estimated the functional connectivity using two different methods: the coherence and the debiased weighted phase lag index (WPLI). The latter metric was designed to reduce the impact of correlated noise such as volume-conduction and to eliminate the sample-size bias (Vinck, Oostenveld, Wingerden, Battaglia, & Pennartz, 2011), whereas the coherence measure keeps the zero-lag phase synchrony. We measured the functional connectivity for both theta and alpha frequency bands, which produced similar results. We therefore only show the patterns obtained from the alpha band (8 – 15 Hz). Figure 3-6 shows the topographic map of the estimated functional connectivity in alpha band during the delay period (the left and middle columns) and the visual period (the right column). The first thing we note is that during delay period, the coherence pattern (the top left) shows a strong neighboring synchrony, which is commonly due to the volume conduction. Once removing the task-irrelevant noise by deducting the baseline connectivity (the top middle), there seems to be a decrease of 87 synchrony between lateral frontal/parietal channels and lateral occipital channels following the introduction of the cue, while most other channels are more synchronized. The WPLI (the bottom left and bottom middle), in comparison, seems to highlight two local networks emerging during the delay period, one in the frontal/parietal region and the other one in the occipital region. These results illustrate two different patterns of functional connectivity underlying attention processing when considering the zero phase lag and non-zero phase lag components, respectively. In addition, the right column shows the difference of functional connectivity patterns between attended and unattended conditions. As we can see from the small range of the color bar, the attention condition does not make a large impact on the connectivity strength, despite a general trend of negative modulation. 88 Figure 3-6. The functional connectivity estimates. Each subplot is a topographic visualization of the functional connectivity estimate under different conditions, with the color and thickness of lines coding the synchrony strength. The top row shows the results using coherence measure, and the bottom row shows results with debiased WPLI. The left column shows the original values of the connectivity estimated during the delay period. The middle column shows the change of connectivity pattern during delay compared with the baseline period. The right column shows the contrasting difference between attended (feature match) and unattended (feature nonmatch) conditions. 3.3 Discussion In summary, we carried out an EEG experiment to explore the behavioral effect of neural synchrony and the global brain network pattern in human, both underlying 89 the feature-based attention. Our preliminary results indicate that human detection performance depends on the phase of alpha-oscillation just following stimulus onset, and interestingly, such phase dependency is suppressed by feature-based attention. Additionally, our functional connectivity analysis revealed distinct task-dependent global network patterns with or without considering the zero-lag phase synchrony. These findings provide a more global perspective on the dynamics of feature based attention and provide complementary insights to the conclusions presented in Chapter 2. Starting from the traditional inhibition-of-return (IOR) studies, it has been well observed that behavioral performance or reaction time may show an oscillatory pattern that is triggered by a task event or directly phase-locked to the recorded oscillatory neural activity. For instance, it was shown that the detection of sensory stimuli is modulated by the phase of the ongoing gamma oscillation in sensory areas, even by optogenetically induced gamma activities (Meletis et al., 2009; Rohenkohl et al., 2018). Additionally, recent work demonstrated a theta-frequency performance rhythm that seemed to reveal the alternating shift of spatial attention focus, which highly depends on the theta oscillation in the frontal-parietal network (Fiebelkorn & Pinsk, 2018; Fiebelkorn et al., 2013). Therefore, the phasic modulation of behavior not only provides a causal indicator for the oscillatory neural dynamics, but may also reveal how neural oscillations impact the perceptual decision-making process underlying various forms of cognitive modulation, such as attention. 90 As we introduced in section 3.1, spatial attention was shown to suppress the theta phase-detection dependency for stimuli that appear at the attended location while the detection of stimuli outside the attended range was still subject to a sampling rhythm (Harris et al., 2018). Consistent with this study, monkey electrophysiological results also showed that the theta-gamma coupling in the ventral stream regions decreased with attention (Spyropoulos, Bosman, & Fries, 2018). These results suggested that spatial attention might serve to disconnect the processing of attended information from a rhythmic sampling of multiple locations, which inspired our experiment. Our preliminary data reveals a similar mechanism for feature-based attention, except that the rhythm for unattended detection is in the alpha band. Strikingly, these results accord with our cross-frequency analysis in the monkey experiment (section 2.2.8), i.e., the reduction of alpha-gamma coupling between channels that shared preference with the attended feature. Combining these results, we believe that the alpha rhythm might act as a feature-sampling mechanism that is modulated by feature-based attention, correspondent with the theta sampling during spatial attention processing. While seemingly promising, we realize that this study suffers from limitation and challenges, mostly due to the lack of topographic organization of feature representation in the visual cortex. In contrast, the visuospatial domain is well structured with retinotopic maps, so that neurophysiological studies could at least take advantage of the contralateral preference and evaluate the difference between the attended and unattended processing by simply contrasting the recordings from 91 the cued side versus the uncued side. In this way, Harris et al., 2018 were able to measure the modulation of power spectrum mediated by spatial attention and found that the theta power was not decreased, but enhanced on the attended side. This result indicated that the reduction of behavioral phase dependency would not be attributed to a decrease of theta power in sensory areas, but rather reflected a modulation at the later-stage decision level. Unfortunately, such a comparison is not applicable in the feature domain with the current spatial resolution of non-invasive recording techniques, which largely limits our understanding of the phase-detection relationship underlying feature-based attention. It would be ideal to repeat the same experiment in animal models using large-scale invasive (high-resolution) recording methods. 3.4 Method 3.4.1 Participants 8 human subjects (4 males and 4 females with age ranging from 24 to 35) participated in this study. All subjects were included in most analyses whereas the phase dependency analysis (section 3.2.4) excluded one subject due to the limited trial number for each condition. All subjects self-reported to have normal or corrected-to- normal vision and normal color vision. One subject was left-handed while the rest reporting being right-handed. Subjects were compensated $50 for their 3.5~4 hours of participation. The study was approved by the Brown University IRB and all procedures followed the approved protocols. 92 3.4.2 Behavioral Paradigm In the cued detection task, a square pattern (16°*16°) of flickering random colored dots was constantly shown in the background. Six hundred ms after the trial began, a bright color cue (a 0.7° dot of either green or purple color) was shown for 200 ms on either left or right side of the fixation spot at an eccentricity of 4° visual angle. Following a variable delay period (400 ms - 650 ms), a target was formed by a round patch (radius = 2.7°) of flickering dots with a consistent color (purple or green) appearing at either the left or right possible location. The subject’s task was to detect the onset of the target and report with button press as soon as a “go” cue (fixation spot turns to a white diamond) was presented. Across all trials, 53.3% were color- matched, i.e., the cue predicted the color of the target, 26.7% were color-nonmatched, and 20% were catch trials, where the target was never presented. The faintness (i.e., the color coherence) of the target was adjusted for the color-matched and color- nonmatched conditions separately to keep the subject’s performance around threshold level performance (set as ~50% detection rate). Throughout each trial, the subject was asked to maintain fixation at the central spot on the screen, and their eye position is monitored using the Eyelink eye tracker with 1000 Hz sampling rate. 3.4.3 Visual stimuli The visual stimuli shown in the task were generated using the STIM (a stimulus generating system) running on Windows 10, and the experimental task was controlled with the ESS state system running on QNX4 real-time operating system, both developed by Dr. David Sheinberg and colleagues. All stimuli were presented on 93 a Display++ LCD monitor (68 cm * 38 cm) at a 100 Hz refresh rate, positioned in front of the subject with a 70 cm viewing distance. The luminance of the gray background matched the average luminance of the flickering dot pattern (23.5 cd/m2). The on/off state and the color of the dots were randomly assigned every frame with an even probability distribution. When a target was shown, the dots within the target area had a higher chance of changing to the target color, and this probability (i.e., color coherence) determined the difficulty of the detection task. 3.4.4 Data acquisition and processing The EEG signal was collected using a 64-channel BrainCap MR electrode cap (Ag/AgCl ring electrode, 10 MΩ) made by Brain Products. Due to hardware limitations, we selectively picked 32 channels to record from (Figure 3-7). The signal was amplified and digitized by the Tucker-Davis Technology (TDT) digitizer PZ5 (at 50 kHz sampling rate) and processed using the TDT data acquisition system RZ5, through which the raw signal was band-passed (1~300 Hz) and down-sampled at a 1017 Hz rate. We did not use a reference channel during the online signal processing, whereas in the offline processing, we re-referenced the signal of each channel to the average across all recorded channels. To reduce the impact of volume conduction and increase the spatial resolution, we then applied a surface Laplacian to the data. All the preprocessing and following analysis were performed using the Python-based toolbox MNE (Gramfort et al., 2013, 2014). 94 Figure 3-7. The channel map of the EEG recording. The cap uses a customized montage. The vertical line represents the anterior- posterior axis, and the horizontal line represents the midline sagittal plane. Channel 59 corresponds to the channel “lz” in the 10-20 system. Recorded channels are marked with yellow circles. 3.4.5 Data analysis Following the pre-processing, we realigned the data to the cue onset and the target onset, respectively, and then removed the outlier trials when the EEG data exceeds 200 μV. In our topographic visualization analysis, we first computed the power time- frequency spectrogram for each channel in each subject using Morlet Wavelets. The 95 number-of-cycle parameter (i.e., the time window size) was determined partially based on the frequency (N_cycle = frequency/6+0.5) in order to resolve the tradeoff between temporal resolution and estimation robustness. The results from each corresponding electrode site were then averaged across subjects, leading to a topographic distribution of power spectrogram. For the phase dependency test, we adopted the method used in Harris et al., 2018 (originally introduced in Vanrullen, 2016), which was based on a permutation test on the phase distribution among conditions. First, we computed the time-frequency representation (complex numbers) of the data for each trial using the same Morlet Wavelet method as above. The trials were grouped by the behavioral response (hit or miss) and the phase distribution was considered using the inter-trial coherence (ITC). Figure 3-8(a) shows example distributions and their ITC, with the left set representing a case of high phase dependency and the right one representing low phase modulation. The assumption is that if the phase has an impact on the behavioral outcome, the phase distribution for the two groups should both show higher clustering (larger ITC) and peaks at different phases. In other words, the summation of ITChit and ITCmiss should be higher than ITCall. Therefore, the phase opposition summation (POS) was defined to describe such phase bias: 𝑃𝑃𝑃𝑃𝑃𝑃 = 𝐼𝐼𝐼𝐼𝐼𝐼ℎ𝑖𝑖𝑖𝑖 + 𝐼𝐼𝐼𝐼𝐼𝐼𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 − 2 ∙ 𝐼𝐼𝐼𝐼𝐼𝐼𝑎𝑎𝑎𝑎𝑎𝑎 Using POS as the measure, we then shuffled the hit/miss assignment on the data 1000 times and compared the original POS values with the permutated results using z- score tests, as depicted in Figure 3-8(b). The p value of this test was then used to guide 96 our choice of the frequency and time window for evaluating the neural phase dependency of behavior. For the phase dependency visualization, we took the phase values within the time- frequency window mentioned above. For each time-frequency point and each focal electrode, we divided the phase among different trials into 7 groups (evenly binned from -π to π) and counted the average hit rate within each group. The one with the highest hit rate was identified as the “best” phase and was aligned in the center with its neighboring bins staying next to it. Such performance-phase function was calculated and averaged over all time-frequency points within the window of interest and all focal electrodes. The same procedure was repeated for the attended trials and unattended trials separately. 97 Figure 3-8. The phase opposition measurement and test. (a) Examples of phase distribution for the hit and miss conditions: higher phase clustering and modulation on the left, and lower phase clustering and modulation on the right. (b) The diagram of the POS permutation test (adapted from Vanrullen, 2016). Data was divided by the hit and miss conditions and the POS pattern was computed for each condition. 98 For the functional connectivity reconstruction, we considered the phase relationship between each pair of signals during the baseline (-300ms–0ms relative to the cue onset), delay (400ms–700ms relative to the cue onset), and visual (300ms–600ms relative to the probe onset) periods. The phase relationship was measured using two methods: the coherence and the debiased weighted phase lag index (WPLI), as they focus on different aspects of the synchrony. Coherence measures how coherently the phase is distributed among trials using: �𝐸𝐸(𝑆𝑆𝑥𝑥𝑥𝑥 )� 𝐶𝐶𝐶𝐶ℎ = �𝐸𝐸(𝑆𝑆𝑥𝑥𝑥𝑥 ) ∙ 𝐸𝐸(𝑆𝑆𝑦𝑦𝑦𝑦 ) where Sxx and Syy are the auto-spectral density for both signals, Sxy stands for their cross-spectral density, and E(∙) represents the average across trials. In contrast, WPLI discards the zero-lag phase component by taking statistics based on the imaginary component of the cross-spectral density: �𝐸𝐸�𝐼𝐼𝐼𝐼�𝑆𝑆𝑥𝑥𝑥𝑥 ��� 𝑊𝑊𝑊𝑊𝑊𝑊𝑊𝑊 = 𝐸𝐸��𝐼𝐼𝐼𝐼�𝑆𝑆𝑥𝑥𝑥𝑥 ��� where Im(∙) represents the imaginary part of the complex-numbered spectral density. The debiased WPLI is an modified estimator for the squared WPLI designed to overcome the sample-size bias (Vinck et al., 2011). We computed the coherence and the debiased WPLI using MNE toolbox. Both were calculated based on the Morlet Wavelet transformation with the same parameters used in our analysis of power spectrum. We took the average of alpha-band (8 – 15 99 Hz) from the spectral results to show the functional connectivity in 0, although the theta-band (3 – 7 Hz) produced similar patterns. 100 Chapter 4 Final remarks This thesis mainly centers on the neural dynamics underlying feature-based attention modulation, starting from a literature review of attention studies to our two projects that investigated feature-based attention from different perspectives. In the main project, we focused on the network-level activities within the intermediate visual area V4, where bottom-up and top-down signals meet and are integrated. Our neurophysiological recordings revealed a feature-dependent interaction pattern among spatially separate neurons, which is further modulated by feature-based attention. Our human EEG project, in comparison, zoomed out to look at the large- scale oscillation of the brain and its behavioral effect. We identified a suppressive role of feature-based attention on how behavioral performance depends on the internal alpha-rhythm phase. Despite the diverse aims, methods and results, our studies have provided complementary insights into the nature of feature attention. One thing we would like to note is a congruent pattern between the attentional modulation identified in both studies. As described in previous chapters, we found in monkey area V4 that a low-frequency synchrony pattern persisted before visual stimuli onset and feature-based attention decreased the alpha-gamma coupling within these networks. Surprisingly, our human study revealed that feature-based attention also decorrelated the behavioral and neural oscillations in the alpha range. These together suggested a mechanism of feature-based attention: while alpha wave seems to dominate earlier visual areas as a default internal state, certain feature- 101 specific neural populations may be desynchronized by the top-down attentional signal, such that it gains a higher signal-to-noise ratio for representing the bottom-up information. The de-entrainment to low-frequency fluctuation has been previously observed in spatial attention studies (Esghaei et al., 2015; Harris et al., 2018), suggesting a shared mechanism underlying both types of attention. The other key finding in our V4 study is the modulation of gamma-band synchronization in feature dimensions, which provides further evidence for a unified mechanism of spatial and feature-based attention. Our results demonstrated a dynamic functional connectivity within feature-specific networks, i.e., the engagement of specific feature modules (either by bottom-up inputs or top-down attentional signals) leads to higher cross-modular gamma synchrony, even for those with large physical or retinotopic distance. This pattern can be treated as a homology to the local gamma synchrony induced by spatial attention (Fries et al., 2001), whereby the representation and transmission of the attended information (spatial or feature) is improved through synchronous neural activity. Therefore, spatial attention and feature-based attention not only share basic mechanisms at the single-neuron level, but also rely on similar elements of network coordination, i.e., the alpha-desynchronization and gamma-coupling among attention-relevant neural sources. One major difference is in the spatial layout of the network interaction patterns. Whereas spatial-domain modulation usually operates on a local range of neural population, feature-based attention produces a large-scale modulation on distributed neural ensembles, as shown in our results. However, such 102 modulation can be treated as “local” if we consider the neural representation in feature maps. That is to say, the dynamic functional connectivity provides such a highway network that distributed neural modules can celebrate the same efficiency as the locally clustered representation within a topographic map, like “wormholes” that connects different modules on the 2-dimensional cortex sheet. Beyond feature-based attention, in fact, we expect that the large-range synchronous coordination may serve as a more general mechanism for top-down modulation in topographically non-specific domains. Future studies along this direction may investigate the inter-modular interaction underlying different top-down control, such as expectation or working memory in a non-spatial manner. It is also worth exploring such effects in other sensory modalities. Moreover, it would be advantageous if recordings can be done simultaneously from both the sensory area and the source area for top-down signals, which could enable conditional causality analyses for differentiating the effect of a common source versus horizontal interactions. These studies can be challenging, however, as we pointed out above that experiments in non-spatial domains could suffer from the limitation of recording methods, as the lack of topographic information would require a higher signal resolution (at single- neuronal level), a larger spatial coverage (across modules with a large sample size), and a fine temporal resolution (for synchrony). This may also explain the limited number of studies on feature-based attention until now. With the rapid development of novel neurophysiological techniques, though, we hope that many mysteries about feature-based attention/top-down control will be uncovered soon. 103 Figure 4-1. Summary of the thesis. Our two projects spanned over species and scales but both focused on the neural dynamics of feature-based attention. We investigated the network interactions in the monkey’s area V4 and identified a feature-tuning-dependent synchrony pattern. The feature-based attention regulates this pattern by increasing the gamma synchrony and decreasing the lower-frequency synchrony as well as the cross-frequency coupling. In addition, we explored the large-scale neural dynamics in humans and its relationship with the behavioral performance. We found that feature-based attention suppresses the alpha phase dependency of behavioral performance. Strikingly these results showed a very similar pattern between the two experiments, which is also consistent with the phenomena found in spatial attention studies. These together suggested a common mechanism of attention modulation, i.e., during the resting or unattended state, a global alpha wave covers the whole visual areas and inhibits the sensory processing in general. When attention is evoked, the top-down signal suppresses the effect of the alpha inhibition to allow certain subnetworks to better work on the bottom-up processing. Among these subnetworks (e.g., the feature modules in V4), a gamma-synchrony can be established during visual 104 processing and enhanced by attention for a higher signal efficacy. Although the feature-based modulation has to operate on the distributed, large-scale system, compared with the local modulation of spatial attention, our results indicate that the brain uses a similar synchrony mechanism to bring these separate clusters closer, just like a wormhole that opens and closes flexibly and connects separate modules to provide higher computation efficiency in the high-dimensional brain. 105 References Ahissar, M., & Hochstein, S. (2004). The reverse hierarchy theory of visual perceptual learning. Trends in Cognitive Sciences, 8(10), 457–464. https://doi.org/10.1016/j.tics.2004.08.011 Arall, M., Romeo, A., & Supèr, H. (2012). Role of Feedforward and Feedback Projections in Figure-Ground Responses. In Visual Cortex - Current Status and Perspectives. https://doi.org/10.5772/47753 Baluch, F., & Itti, L. (2011). Mechanisms of top-down attention. Trends in Neurosciences, 34(4), 210–224. https://doi.org/10.1016/j.tins.2011.02.003 Bichot, N. P., Rossi, A. F., & Desimone, R. (2005). Parallel and serial neural mechanisms for visual search in macaque area V4. Science, 308(5721), 529–534. https://doi.org/10.1126/science.1109676 Bisley, J. W., & Goldberg, M. E. (2010). Attention , Intention , and Priority in the Parietal Lobe. Annual Review of Neuroscience. https://doi.org/10.1146/annurev-neuro-060909- 152823 Bollimunta, A., Mo, J., Schroeder, C. E., & Ding, M. (2011). Neuronal Mechanisms and Attentional Modulation of Corticothalamic Alpha Oscillations. The Journal of Neuroscience, 31(13), 4935–4943. https://doi.org/10.1523/JNEUROSCI.5580-10.2011 Brooks, D. I., & Sheinberg, D. L. (2015). Modified toolbox for optogenetics in the nonhuman primate. Neurophotonics. https://doi.org/10.1117/1.NPh.2.3.031202 Bruce, C. J., & Goldberg, M. E. (1985). Primate Frontal Eye Fields . I . Single Neurons Discharging Before Saccades. Journal of Neurophysiology, 53(3). https://doi.org/10.1152/jn.1985.53.3.603 Buffalo, E. a, Fries, P., Landman, R., Buschman, T. J., & Desimone, R. (2011). Laminar differences in gamma and alpha coherence in the ventral stream. Proceedings of the National Academy of Sciences of the United States of America, 108(27), 11262–11267. https://doi.org/10.1073/pnas.1011284108 Buffalo, E. a, Fries, P., Landman, R., Liang, H., & Desimone, R. (2010). A backward progression of attentional effects in the ventral stream. Proceedings of the National Academy of Sciences of the United States of America, 107(1), 361–365. https://doi.org/10.1073/pnas.0907658106 Buschman, T. J., Denovellis, E. L., Diogo, C., Bullock, D., & Miller, E. K. (2012). Synchronous Oscillatory Neural Ensembles for Rules in the Prefrontal Cortex. Neuron, 76(4), 838–846. https://doi.org/10.1016/j.neuron.2012.09.029 Buschman, T. J., & Kastner, S. (2015). From Behavior to Neural Dynamics: An Integrated Theory of Attention. Neuron, 88(1), 127–144. https://doi.org/10.1016/j.neuron.2015.09.017 106 Buschman, T. J., & Miller, E. K. (2007). Top-down versus bottom-up control of attention in the prefrontal and posterior parietal cortices. Science, 315(8520), 1860–1863. https://doi.org/10.1126/science.1138071 Camperi, M., & Wang, X. (1998). A Model of Visuospatial Working Memory in Prefrontal Cortex : Recurrent Network and Cellular Bistability, 405, 383–405. Canolty, R. T., & Knight, R. T. (2010). The functional role of cross-frequency coupling. Trends in Cognitive Sciences, 14(11), 506–515. https://doi.org/10.1016/j.tics.2010.09.001 Carrasco, M., Ling, S., & Read, S. (2004). Spatial attention alters visual appearance. Nature Neuroscience, 7(3), 308–313. https://doi.org/10.1038/nn1194 Chun, M. M., & Turk-browne, N. B. (2007). Interactions between attention and memory. Current Opinion in Neurobiology, 17(2). https://doi.org/10.1016/j.conb.2007.03.005 Clark, K. L., Noudoost, B., & Moore, T. (2012). Persistent Spatial Information in the Frontal Eye Field during Object-Based Short-Term Memory. The Journal of Neuroscience, 32(32), 10907–10914. https://doi.org/10.1523/JNEUROSCI.1450-12.2012 Clark, K., & Noudoost, B. (2018). Concurrent influence of top-down and bottom-up inputs on correlated activity of Macaque extrastriate neurons. Nature Communications. https://doi.org/10.1038/s41467-018-07816-4 Compte, A., & Wang, X.-J. (2006). Tuning curve shift by attention modulation in cortical neurons: a computational study of its mechanisms. Cerebral Cortex, 16(6), 761–778. https://doi.org/10.1093/cercor/bhj021 Connor, C. E., Gallant, J. L., Preddie, D. C., & Van Essen, D. C. (1996). Responses in area V4 depend on the spatial relationship between stimulus and attention. Journal of Neurophysiology, 75(3), 1306–1308. Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/8867139 Craighero, L., & Rizzolatti, G. (2005). CHAPTER 31 - The Premotor Theory of Attention. In L. Itti, G. Rees, & J. K. B. T.-N. of A. Tsotsos (Eds.) (pp. 181–186). Burlington: Academic Press. https://doi.org/https://doi.org/10.1016/B978-012375731-9/50035-5 David, S. V, Hayden, B. Y., Mazer, J. a, & Gallant, J. L. (2008). Attention to stimulus features shifts spectral tuning of V4 neurons during natural vision. Neuron, 59(3), 509–521. https://doi.org/10.1016/j.neuron.2008.07.001 Egner, T., Monti, J. M. P., Trittschuh, E. H., Wieneke, C. A., Hirsch, J., & Mesulam, M.-M. M. (2008). Neural integration of top-down spatial and feature-based information in visual search. The Journal of Neuroscience, 28(24), 6141–6151. https://doi.org/10.1523/JNEUROSCI.1262-08.2008 Esghaei, M., Daliri, M. R., & Treue, S. (2015). Attention Decreases Phase-Amplitude Coupling, Enhancing Stimulus Discriminability in Cortical Area MT. Frontiers in Neural Circuits, 9(December). https://doi.org/10.3389/fncir.2015.00082 Essen, D. C. Van. (2005). Role of Primate Visual Area V4 in the Processing of 3-D Shape Characteristics Defined by Disparity. Journal of Neurophysiology, 94(4), 2856–2866. https://doi.org/10.1152/jn.00802.2004 Fiebelkorn, I. C., & Pinsk, M. A. (2018). A Dynamic Interplay within the Frontoparietal A 107 Dynamic Interplay within the Frontoparietal Network Underlies Rhythmic Spatial Attention. Neuron, 99(4), 842–853.e8. https://doi.org/10.1016/j.neuron.2018.07.038 Fiebelkorn, I. C., Pinsk, M. A., & Kastner, S. (2019). The mediodorsal pulvinar coordinates the macaque fronto-parietal network during rhythmic spatial attention. Nature Communications, 10(215). https://doi.org/10.1038/s41467-018-08151-4 Fiebelkorn, I. C., Saalmann, Y. B., & Kastner, S. (2013). Rhythmic sampling within and between objects despite sustained attention at a cued location. Current Biology, 23(24), 2553– 2558. https://doi.org/10.1016/j.cub.2013.10.063 Fougnie, D. (2008). The relationship between attention and working memory. New Research on Short-Term Memory, 1–45. Retrieved from http://visionlab.harvard.edu/Members/darylfougnie/Daryl_Fougnie_(Academic)/Ho me_files/Fougnie-in press-chap 1.pdf Fries, P. (2005). A mechanism for cognitive dynamics: neuronal communication through neuronal coherence. Trends in Cognitive Sciences, 9(10), 474–480. https://doi.org/10.1016/j.tics.2005.08.011 Fries, P. (2009). Neuronal Gamma-Band Synchronization as a Fundamental Process in Cortical Computation. Annual Review of Neuroscience. https://doi.org/10.1146/annurev.neuro.051508.135603 Fries, P., Reynolds, J. H., Rorie, A. E., & Desimone, R. (2001). Modulation of Oscillatory Neuronal Synchronization by Selective Visual Attention. Science, 291(February), 1560– 1564. Fries, P., Womelsdorf, T., Oostenveld, R., & Desimone, R. (2008). The Effects of Visual Stimulation and Selective Visual Attention on Rhythmic Neuronal Synchronization in Macaque Area V4. The Journal of Neuroscience, 28(18), 4823–4835. https://doi.org/10.1523/JNEUROSCI.4499-07.2008 Fuster, J. M., & Alexander, G. E. (1971). Neuron Activity Related to Short-Term Memory. Science, 173(3997), 652–654. Gold, J. I., & Shadlen, M. N. (2007). The Neural Basis of Decision Making. Annual Review of Neuroscience. https://doi.org/10.1146/annurev.neuro.29.051605.113038 Gottlieb, J. P., Kusunoki, M., & Goldberg, M. E. (1998). The representation of visual salience in monkey parietal cortex. Nature, 391, 481–484. https://doi.org/10.1038/35135 Gramfort, A., Luessi, M., Larson, E., Engemann, D. A., Strohmeier, D., Brodbeck, C., … Hämäläinen, M. (2013). MEG and EEG data analysis with MNE-Python. Frontiers in Neuroscience, 7, 267. https://doi.org/10.3389/fnins.2013.00267 Gramfort, A., Luessi, M., Larson, E., Engemann, D. A., Strohmeier, D., Brodbeck, C., … Hämäläinen, M. S. (2014). NeuroImage MNE software for processing MEG and EEG data. NeuroImage, 86, 446–460. https://doi.org/10.1016/j.neuroimage.2013.10.027 Gray, C. M., Konig, P., Engel, A. K., & Singer, W. (1989). Oscillatory responses in cat visual cortex exhibit inter-columnar global stimulus properties. Nature, 338(March). Gregoriou, G. G., Gotts, S. J., Zhou, H., & Desimone, R. (2009). High-Frequency, Long-Range Coupling Between Prefrontal and Visual Cortex During Attention. Science, 324(May), 108 1207–1211. Grothe, I., Neitzel, S. D., Mandon, S., & Kreiter, A. K. (2011). Attention induces dynamic changes in coherence between monkey areas V1 and V4. Nature Precedings. https://doi.org/10.1038/npre.2011.5884.1 Harris, A. M., Dux, P. E., & Mattingley, J. B. (2018). Detecting unattended stimuli depends on the phase of pre-stimulus neural oscillations. The Journal of Neuroscience, 38(12), 3006– 3017. https://doi.org/10.1523/JNEUROSCI.3006-17.2018 Herrmann, K., Montaser-Kouhsari, L., Carrasco, M., & Heeger, D. J. (2010). When size matters: attention affects performance by contrast or response gain. Nature Neuroscience, 13(12), 1554–1559. https://doi.org/10.1038/nn.2669 Hillis, A. E., Newhart, M., Heidler, J., Barker, P. B., & Herskovits, E. H. (2005). Anatomy of Spatial Attention : Insights from Perfusion Imaging and Hemispatial Neglect in Acute Stroke. The Journal of Neuroscience, 25(12), 3161–3167. https://doi.org/10.1523/JNEUROSCI.4468-04.2005 Hoogenboom, N., Schoffelen, J., Oostenveld, R., & Fries, P. (2010). NeuroImage Visually induced gamma-band activity predicts speed of change detection in humans. NeuroImage, 51(3), 1162–1167. https://doi.org/10.1016/j.neuroimage.2010.03.041 Kajikawa, Y., & Schroeder, C. E. (2011). How local is the local field potential? Neuron, 72(5), 847–858. https://doi.org/10.1016/j.neuron.2011.09.029 Leavitt, M. L., Mendoza-halliday, D., & Martinez-trujillo, J. C. (2017). Sustained Activity Encoding Working Memories : Not Fully Distributed. Trends in Neurosciences, 40(6), 328–346. Li, M., Liu, F., Juusola, M., & Tang, S. (2014). Perceptual Color Map in Macaque Visual Area V4. The Journal of Neuroscience, 34(1), 202–217. https://doi.org/10.1523/JNEUROSCI.4549-12.2014 Ling, S., & Carrasco, M. (2005). Contrast gain vs. response gain: Do sustained and transient covert attention exhibit different signature responses? Journal of Vision, 5(8). https://doi.org/10.1167/5.8.686 Luck, S., & Chelazzi, L. (1997). Neural mechanisms of spatial selective attention in areas V1, V2, and V4 of macaque visual cortex. Journal of Neurophysiology, 24–42. Retrieved from http://www.atahiri.com/dev/desimone/wp-content/uploads/2013/11/Neural- mechanisms-of-spatial-selective-attention-in-areas-V1-V2-and-V4-of-macaque-visual- cortex_Journal-of-Neurophysiology_1997.pdf Luck, S. J., Chelazzi, L., Hillyard, S. A., Desimone, R., Diego, S., Jolla, L., … Hillyard, S. A. (1997). Neural Mechanisms of Spatial Selective Attention in Areas V1 , V2 , and V4 of Macaque Visual Cortex. Journal of Neurop, 77(1), 24–42. https://doi.org/10.1152/jn.1997.77.1.24 Lundqvist, M., Herman, P., & Miller, X. E. K. (2018). Working Memory: Delay Activity , Yes ! Persistent Activity ? Maybe Not. The Journal of Neuroscience, 38(32), 7013–7019. https://doi.org/10.1523/JNEUROSCI.2485-17.2018 McAdams, C. J., & Maunsell, J. H. (1999). Effects of attention on orientation-tuning functions of single neurons in macaque cortical area V4. The Journal of Neuroscience, 19(1), 431– 441. Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/9870971 109 Meldolesi, J., Kirchhoff, F., Navarrete, M., Araque, A., Jahr, C. E., Rubio, M. E., … Rhode, D. (2012). The Pulvinar Regulates Information Transmission Between Cortical Areas Based on Attention Demands. Science, (August), 753–757. Meletis, K., Knoblich, U., Zhang, F., Deisseroth, K., Cardin, J. A., Carle, M., … Tsai, L. (2009). Driving fast-spiking cells induces gamma rhythm and controls sensory responses. Nature, 459(June), 663–668. https://doi.org/10.1038/nature08002 Mendoza-halliday, D., Torres, S., & Martinez-trujillo, J. C. (2014). Sharp emergence of feature- selective sustained activity along the dorsal visual pathway. Nature Publishing Group, 17(9), 1255–1262. https://doi.org/10.1038/nn.3785 Merrikhi, Y., Clark, K., Albarran, E., Parsa, M., Zirnsak, M., Moore, T., & Noudoost, B. (2017). Spatial working memory alters the efficac of input to visual cortex. Nature Communications, 8, 1–10. https://doi.org/10.1038/ncomms15041 Miller, E. K., & Buschman, T. J. (2012). Cortical circuits for the control of attention. Current Opinion in Neurobiology, 23(2), 216–222. https://doi.org/10.1016/j.conb.2012.11.011 Mitchell, J. F., Sundberg, K. A., & Reynolds, J. H. (2009). Article Spatial Attention Decorrelates Intrinsic Activity Fluctuations in Macaque Area V4 SHUFFLE. Neuron, 63(6), 879–888. https://doi.org/10.1016/j.neuron.2009.09.013 Moran, J., & Desimone, R. (1985). Selective attention gates visual processing in the extrastriate cortex. Science. Retrieved from http://www.sciencemag.org/content/229/4715/782.short Oostenveld, R., Fries, P., Maris, E., & Schoffelen, J.-M. (2011). FieldTrip: Open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data. Computational Intelligence and Neuroscience, 2011, 156869. https://doi.org/10.1155/2011/156869 Pasupathy, A., & Connor, C. E. (1999). Responses to Contour Features in Macaque Area V4. Journal of Neurophysiology, 82(5), 2490–2502. https://doi.org/10.1152/jn.1999.82.5.2490 Pelt, S. Van, Schoffelen, J., Kennedy, H., Fries, P., Michalareas, G., Vezoli, J., … Kennedy, H. (2016). Alpha-Beta and Gamma Rhythms Subserve Feedback and Feedforward Influences among Human Visual Cortical Areas Article Alpha-Beta and Gamma Rhythms Subserve Feedback and Feedforward Influences among Human Visual Cortical Areas. Neuron, 89(2), 384–397. https://doi.org/10.1016/j.neuron.2015.12.018 Ponce-alvarez, A., Thiele, A., Albright, T. D., Stoner, G. R., & Deco, G. (2013). Stimulus- dependent variability and noise correlations in cortical MT neurons. Proceedings of the National Academy of Sciences, 110(32). https://doi.org/10.1073/pnas.1300098110/- /DCSupplemental.www.pnas.org/cgi/doi/10.1073/pnas.1300098110 Popov, X. T., Kastner, S., & Jensen, X. O. (2017). FEF-Controlled Alpha Delay Activity Precedes Stimulus-Induced Gamma-Band Activity in Visual Cortex. The Journal of Neuroscience, 37(15), 4117–4127. https://doi.org/10.1523/JNEUROSCI.3015-16.2017 Ptak, R. (2012). The frontoparietal attention network of the human brain: action, saliency, and a priority map of the environment. Neuroscientist, 18(5), 502–515. https://doi.org/10.1177/1073858411409051 110 Reichert, D. P., & Serre, T. (2013). Neuronal Synchrony in Complex-Valued Deep Networks. ArXiv. Retrieved from http://arxiv.org/abs/1312.6115 Reynolds, J. H., & Heeger, D. J. (2009). The normalization model of attention. Neuron, 61(2), 168–185. https://doi.org/10.1016/j.neuron.2009.01.002 Roe, A. W., Chelazzi, L., Connor, C. E., Conway, B. R., Fujita, I., Gallant, J. L., … Vanduffel, W. (2012). Toward a unified theory of visual area V4. Neuron, 74(1), 12–29. https://doi.org/10.1016/j.neuron.2012.03.011 Roelfsema, P. R., Lamme, V. a F., Spekreijse, H., & Bosch, H. (2002). Figure-ground segregation in a recurrent network architecture. Journal of Cognitive Neuroscience, 14(4), 525–537. https://doi.org/10.1162/08989290260045756 Rohenkohl, G., Bosman, C. A., & Fries, P. (2018). Improves Behavioral Performance Article Gamma Synchronization between V1 and V4 Improves Behavioral Performance. Neuron, 100(4), 953–963.e3. https://doi.org/10.1016/j.neuron.2018.09.019 Rolls, E. T. (2012). Invariant Visual Object and Face Recognition: Neural and Computational Bases, and a Model, VisNet. Frontiers in Computational Neuroscience, 6(June), 35. https://doi.org/10.3389/fncom.2012.00035 Ruff, D. A., & Cohen, M. R. (2014). Attention can either increase or decrease spike count correlations in visual cortex. Nature Neuroscience, 17(11), 1591–1598. https://doi.org/10.1038/nn.3835 Schwabe, L., & Obermayer, K. (2005). Adaptivity of tuning functions in a generic recurrent network model of a cortical hypercolumn. The Journal of Neuroscience : The Official Journal of the Society for Neuroscience, 25(13), 3323–3332. https://doi.org/10.1523/JNEUROSCI.4493-04.2005 Siegle, J. H., Pritchett, D. L., & Moore, C. I. (2014). Gamma-range synchronization of fast- spiking interneurons can enhance detection of tactile stimuli. Nature Neuroscience, 17(10), 1371–1379. https://doi.org/10.1038/nn.3797 Simons, D. J., & Chabris, C. F. (1999). Gorillas in our midst: sustained inattentional blindness for dynamic events. Perception, 28(9), 1059–1074. https://doi.org/10.1068/p281059 Spyropoulos, G., Arturo, C., & Fries, P. (2018). A theta rhythm in macaque visual cortex and its attentional modulation. Proceedings of the National Academy of Sciences, 115(24). https://doi.org/10.1073/pnas.1719433115 Spyropoulos, G., Bosman, C. A., & Fries, P. (2018). A theta rhythm in macaque visual cortex and its attentional modulation. Proceedings of the National Academy of Sciences, 115(24), E5614–E5623. https://doi.org/10.1073/pnas.1719433115 Störmer, V. S., & Alvarez, G. A. (2014). Feature-Based Attention Elicits Surround Suppression in Feature Space. Current Biology, 30(19), 1985–1988. https://doi.org/10.1016/j.cub.2014.07.030 Summerfield, C., & Egner, T. (2016). Feature-Based Attention and Feature-Based Expectation. Trends in Cognitive Sciences, 20(6), 401–404. https://doi.org/10.1016/j.tics.2016.03.008 Tiesinga, P. H., Fellous, J., Salinas, E., Jose, J. V, & Sejnowski, T. J. (2005). Inhibitory synchrony 111 as a mechanism for attentional gain modulation. Journal of Physiology, Paris, 98(2004), 296–314. https://doi.org/10.1016/j.jphysparis.2005.09.002 Treisman, A. (1998). Feature binding, attention and object perception. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences, 353(1373), 1295–1306. https://doi.org/10.1098/rstb.1998.0284 Tsuchiya, N., & Koch, C. (2008). Attention and consciousness. Scholarpedia. https://doi.org/10.4249/scholarpedia.4173 Ungerleider, L. G., Galkin, T. W., Desimone, R., & Gattass, R. (2008). Cortical connections of area V4 in the macaque. Cerebral Cortex, 18(3), 477–499. https://doi.org/10.1093/cercor/bhm061 van Kerkoerle, T., Self, M. W., Dagnino, B., Gariel-Mathis, M. -a., Poort, J., van der Togt, C., & Roelfsema, P. R. (2014). Alpha and gamma oscillations characterize feedback and feedforward processing in monkey visual cortex. Proceedings of the National Academy of Sciences, 111(40). https://doi.org/10.1073/pnas.1402773111 Vanrullen, R. (2016). How to Evaluate Phase Differences between Trial Groups in Ongoing Electrophysiological Signals. Frontiers in Neuroscience, 10(September), 1–22. https://doi.org/10.3389/fnins.2016.00426 Vinck, M., Oostenveld, R., Wingerden, M. Van, Battaglia, F., & Pennartz, C. M. A. (2011). NeuroImage An improved index of phase-synchronization for electrophysiological data in the presence of volume-conduction , noise and sample-size bias. NeuroImage, 55(4), 1548–1565. https://doi.org/10.1016/j.neuroimage.2011.01.055 Vinck, M., van Wingerden, M., Womelsdorf, T., Fries, P., & Pennartz, C. M. A. (2010). The pairwise phase consistency: A bias-free measure of rhythmic neuronal synchronization. NeuroImage, 1–11. https://doi.org/10.1016/j.neuroimage.2010.01.073 Wang, X. (2002). Probabilistic decision making by slow reverberation in cortical circuits, 36, 955–968. Wang, X., & Buzsaki, G. (2012). Mechanisms of Gamma Oscillations. Annual Review of Neuroscience. https://doi.org/10.1146/annurev-neuro-062111-150444 Woloszyn, L., & Sheinberg, D. L. (2009). Neural Dynamics in Inferior Temporal Cortex during a Visual Working Memory Task. The Journal of Neuroscience, 29(17), 5494–5507. https://doi.org/10.1523/JNEUROSCI.5785-08.2009 Womelsdorf, T., & Fries, P. (2007). The role of neuronal synchronization in selective attention. Current Opinion in Neurobiology, 17(2), 154–160. https://doi.org/10.1016/j.conb.2007.02.002 Yoshioka, T., Levitt, J., & Lund, J. (1992). Intrinsic lattice connections of macaque monkey visual cortical area V4. The Journal of Neuroscience, 12(July). Retrieved from http://www.jneurosci.org/content/12/7/2785.short Zhang, M., & Barash, S. (2003). Persistent LIP Activity in Memory Antisaccades : Working Memory For a Sensorimotor Transformation. Journal of Neurophysiology, 1424–1441. https://doi.org/10.1152/jn.00504.2003 Zhou, H., & Desimone, R. (2011). Feature-based attention in the frontal eye field and area V4 112 during visual search. Neuron, 70(6), 1205–1217. https://doi.org/10.1016/j.neuron.2011.04.032 Zhou, H., Schafer, R. J., & Desimone, R. (2016). Article Pulvinar-Cortex Interactions in Vision and Attention. Neuron, 89(1), 209–220. https://doi.org/10.1016/j.neuron.2015.11.034 113