Sequence learning and experience-driven modifications create sparse predictive representations of visual information in mouse primary visual cortex
OA Version
Citation
Abstract
Vision is a widely studied aspect of brain function, and a great deal of research has uncovered the principles by which the cortical hierarchy converts spatiotemporal light patterns into unique neural representations that are the basis of our ability to perceive objects in the visual scene. More recently, primary visual areas have been used to study common principles of cortical function that are shared with other brain regions. Predictive coding is an influential theory of cortical function that might explain some properties of extra classical receptive fields in the visual system but has not been experimentally verified in a general sense. Particularly absent is an understanding of how evoked neural activity modifies cortical circuits to form predictive representations. The visual cortex is highly plastic and can encode spatiotemporal relationships, making it a useful model system to test how experience modifies cortical circuits to encode memories that include temporal information, and how these modification affect the processing of subsequent sensory inputs and are used to make predictions. We begin with a review of visual neuroscience with a particular focus on what is known about the architecture and function of the mouse visual system. We discuss the anatomical circuits that convert light into neural activity and how it this filtered, processed, and refined in the early visual system. This includes a discussion of the coding strategies employed by the retina, properties of thalamic relay cells, and an overview of neocortical circuitry. This section concludes with a review of visual cortical plasticity. Next, we review the theoretical and conceptual framework that underlies predictive processing theories of cortical function. We compare several predictive coding models and discuss implementation details proposed by these theories with respect to the neocortical circuit. We then conduct a review of previous experimental work aimed at testing the extent to which cortical dynamics can be understood as predictive processing. This includes computational studies of visual cortical responses, mismatch negativity, sensorimotor mismatch signals, and recent discoveries of prediction error cells in the visual and auditory systems. We then present our original research investigating predictive and temporal processing in the mouse primary visual cortex (V1). We conducted two-photon calcium imaging in layer 2/3 of the mouse V1 while animals were exposed to visual sequences over several days. We analyzed thousands of cells in search of prediction error responses and characterized how visual experience and expectation change temporally coordinated activity in visually responsive cells. Consistent with predictive coding models, we find that neural activity during following the omission of a predicted element is elevated at the time of expected visual transitions in a trained but not naïve animals. Substituting an unexpected stimulus, however, did not generate significant prediction error responses. Using linear decoders, we show that training has little effect on the ability to accurately decode stimulus identity or time within an experiment despite finding that evoked response dynamics are significantly modified over training. We also show that individual cells’ peak firing times span the temporal duration of both active visual stimulation and interstimulus gray periods. Finally, we show that low-dimensional representations of neural activity become more selective for individual elements of the trained sequence while population level responses become sparser and decorrelated, consistent with the efficient coding hypothesis.
Description
2024
License
Attribution-ShareAlike 4.0 International