Higher-level spatial prediction during natural scene perception in mouse primary visual cortex
Micha Heilbron, Floris de Lange, Donders Institute for Brain, Cognition, and Behaviour, Radboud University, Netherlands
Session:
Posters 3B Poster
Presentation Time:
Sat, 26 Aug, 13:00 - 15:00 United Kingdom Time
Abstract:
Predictive processing models postulate that perception involves the comparison of bottom-up signals with top- down (reconstructive) predictions. Building on recent work, we test this idea in natural scene perception. We use a reconstructing auto-encoder (UNet) to quantify the spatial predictability of natural image patches, and com- pare those predictability estimates to brain responses in the mouse visual system, using a large scale survey of high-density extra-cellular recordings provided by the Allen Institute Brain Observatory. Initial results in V1 re- veal a clear modulation of spiking activity by high (but not low) level visual predictability, suggesting V1 is sensitive to low-level visual features, but high-level predictability. This sensitivity bears striking similarities to recent, self-supervised predictive objectives from machine learning.