Sensory cortex is optimized for prediction of future input

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Neurons in the sensory cortex exhibit selectivity for diverse features in natural scenes.
• The underlying principles determining this neuronal selectivity remain incompletely understood.

Purpose of the Study:

• To investigate whether neuronal selectivity is optimized for representing features that predict immediate future sensory inputs.
• To test the hypothesis that predictive coding shapes receptive field properties in the sensory cortex.

Main Methods:

• Utilized simple feedforward neural networks trained on natural scene video and audio clips.
• Trained networks to predict subsequent moments in sensory data streams.
• Compared network-generated receptive fields with those of real cortical neurons across mammalian species.

Main Results:

• Developed receptive fields in neural networks closely matched real cortical neurons.
• Observed oriented spatial tuning (visual cortex) and frequency selectivity (auditory cortex).
• Found that superior prediction accuracy correlated with more brain-like receptive fields, particularly temporal tuning.

Conclusions:

• Neuronal selectivity in the sensory cortex is shaped by the need to predict future sensory inputs.
• The brain optimizes sensory processing to extract features with high predictive power.
• Predictive coding offers a unifying framework for understanding sensory cortical organization.