Description
- Abstract:
- Biological brains are dynamic. Recent advances in electrophysiology and neuroimaging have helped uncover various mechanisms through which brains construct and utilize rich variations in neural activity patterns for computations. In particular, the primate visual cortex consists of more than thirty densely interconnected areas whose constituent neurons exhibit a wide range of complex activation patterns in response to visual stimulation. Despite this, models of visual function have historically embraced a highly simplified paradigm within which computations are carried out as a finite cascade of operations in serial. The deep feedforward neural network (DNN) family is the modern-day manifestation of such a model class. In recent years, significant technological advances in computing infrastructure coupled with the availability of large, annotated datasets have propelled the adoption of DNNs as the de facto model of vision. DNNs have indeed demonstrated immense proficiency when trained on large-scale, naturalistic visual challenge benchmarks to the extent that they are now on par with, or sometimes beyond, human-level performance. Their impressive successes, however, mask fundamental deficiencies. Our experiments, probing the ability of DNNs to systematically generalize to held-out parameterizations of Pathfinder, a visual challenge task suite, reveal one such deficiency: DNNs can appear to learn fundamental and reusable visual computations while crudely approximating an input-output mapping pertaining to the dataset and task at hand. We posit that this pathology is a consequence of having only a fixed computational budget specified by the number of operations in a model's computational graph. Recurrently connected networks, like those in the primate visual cortex, have the ability to leverage an infinite computational budget through feedback cycles and are promising candidate solutions for the problem mentioned above. However, recurrent network models are hard to train due to well-documented gradient and memory issues. We draw inspiration from neuroscience and derive a novel learning algorithm called Contractive Recurrent Backpropagation (C-RBP) that relies on our model constructing fixed-point attractors. We demonstrate that recurrent vision models trained with C-RBP can not only learn hard parameterizations of Pathfinder but can also successfully generalize to held-out parameterizations at a fraction of parameter- and memory costs when compared to DNNs. We subsequently extend these principles to recurrent vision models trained to solve a large-scale visual challenge (MS-COCO Panoptic Segmentation) and show that our approach outperforms the leading feedforward approach with far fewer free parameters. Furthermore, we show that the visual representation dynamics emerging from our C-RBP trained recurrent neural networks can ably support downstream utilization for decision-making, provide a framework to study the interactions between vision and higher cognitive functions, and facilitate the comparison between model and human behavior in various scenarios.
- Notes:
- Thesis (Ph. D.)--Brown University, 2023
Citation
Govindarajan, Lakshmi Narasimhan,
"Attractor Dynamics in Large Scale Recurrent Neural Networks"
(2023).
Cognitive, Linguistic, and Psychological Sciences Theses and Dissertations.
Brown Digital Repository. Brown University Library.
https://repository.library.brown.edu/studio/item/bdr:ay4c9rbm/
Relations
Collection:
-
Cognitive, Linguistic, and Psychological Sciences Theses and Dissertations
Theses and Dissertations for the Cognitive, Linguistic, and Psychological Sciences department....