It has been established that the feed-forward sweep along the ventral visual pathway is crucial for core object recognition. However, the visual system contains abundant recurrent circuits formed by feedback and horizontal projections, whose functional role in visual perception is largely unknown. My thesis study explored the physiological and computational roles of the recurrent processing during visual perception in a naturalistic viewing scenario. Experimentally, we recorded multi-channel neural activities from areas V4 and IT of non-human primates, while natural images were presented in a passive viewing paradigm that does not require explicit task-dependent cognitive control. To engage recurrent processing mechanisms, we challenged the visual system by introducing occluding noise to images and controlling the image familiarity. Our physiological results showed that in area IT, especially its superficial layers, but not area V4, noise degradation delayed the temporal dynamics of the visual activity and amplified the cross-trial response variability. Moreover, the coded information about visual stimuli was impeded during the early epoch of activity in both brain areas, whereas such interference recovered during the late phase of activity in area IT but not V4. In addition, stimulus familiarity suppressed the neural response posterior to the visual transient, and increased the coding robustness to degradation. We further tested the computational role of recurrent processing with a simple recurrent network model - restricted Boltzmann machine. Even without any parameter tuned to match the neural data, the model successfully replicated the major experimental results, including the delayed response dynamics, increased the cross-trial variability and the recovery of coding robustness over time, all of which emerged through recurrent information integration over space and time in the network. The agreement of neural data and modeling results suggested that the physiological signatures observed in our experiments are the natural corollary of recurrent processing, further supporting that area IT recruited recurrent mechanisms to process visual inputs in challenging scenarios, which increased the robustness of neural code.
All rights reserved. Collection is open to the Brown community for research.
Citation
Guan, Shaobo,
"Vision beyond the feed-forward sweep"
(2019).
Neuroscience Theses and Dissertations.
Brown Digital Repository. Brown University Library.
https://doi.org/10.26300/55qk-0b12