Search Papers | Poster Sessions | All Posters
Poster C21 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Representational sparsity determines representational stability in sensory cortices
Shanshan Qin1,2,3, Cengiz Pehlevan1,2; 1John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, MA 02138, USA., 2Center for Brain Science, Harvard University, Cambridge, MA 02138, USA., 3Center for Computational Neuroscience, Flatiron Institute, New York, NY 10010, USA.
Recent advancements in large-scale neural activity recordings have revealed a continuous evolution in neural population activity associated with familiar tasks, percepts, and actions over extended periods. The underlying mechanisms and functional implications of such "representational drift" remain poorly understood. In many sensory cortices, representation stability varies with stimulus type. For example, in the mouse primary visual cortex, natural movie stimuli induce drift, unlike drifting gratings. To understand the mechanism behind such stimulus-dependent representational drift in visual cortex, we propose that natural stimuli prompt denser responses compared to artificial ones, making denser representations more susceptible to synaptic noise. We evaluated this hypothesis by training a sparse coding network with continually updating synaptic weights. We found that representations for more complex image patches are denser and also exhibit more drift compared to simpler ones. This result is consistent with experimental findings. To further explore the relationship between drift speed and representational sparsity, we developed a mean-field model to analyze how different noise sources contribute to drift. Our model provides a plausible explanation for stimulus-dependent representational drift.
Keywords: Representational drift sparsity visual cortex