Higher-level spatial prediction in natural vision across mouse visual cortex

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Predictive processing theories suggest sensory systems anticipate incoming signals based on context.
• Existing evidence for prediction in sensory cortex often relies on artificial stimuli, not natural perception.

Purpose of the Study:

• To investigate sensory prediction during natural scene perception in the mouse visual cortex.
• To quantify spatial predictability in natural images and correlate it with neural responses.

Main Methods:

• Utilized deep generative modeling to assess spatial predictability of image patches.
• Analyzed large-scale, high-density recordings from the Allen Institute Brain Observatory.
• Controlled for tuning to low-level image features and local statistical context.

Main Results:

• Cortical responses are modulated by sensory predictability; predictable patches elicit weaker responses.
• Visual cortical neurons show sensitivity to higher-level feature predictability, even in primary visual areas.
• Unpredictability sensitivity is more pronounced in superficial layers of the primary visual cortex.
• Spatial prediction effects are independent of recent experience, suggesting reliance on long-term priors.

Conclusions:

• The visual cortex predominantly predicts sensory information at higher levels of abstraction.
• Findings align with predictive coding models and self-supervised learning in artificial intelligence.