Mismatched decoding in the brain.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Information Theory

Background:

• Understanding how the brain decodes information is a fundamental challenge.
• Neural responses exhibit complex correlations that may influence information processing.

Purpose of the Study:

• To develop a framework for simplifying neural decoding models.
• To quantify information loss when using simplified models with mismatched decoders.
• To evaluate the role of neural correlations in information decoding.

Main Methods:

• Constructed hierarchical probabilistic models of neural responses using the maximum entropy principle, ignoring correlations beyond Kth order.
• Introduced an information-theoretic quantity, I*, to evaluate information obtained by mismatched decoders.
• Applied the framework to spike data from the vertebrate retina using natural scene movies as stimuli.

Main Results:

• The proposed quantity I* aligns with minimum mean-square error and mutual information, offering a robust measure.
• A previously proposed measure for correlation importance deviates from I* with large cell populations.
• Despite significant correlations in retinal ganglion cell activity, information loss was negligible when ignoring all correlations during decoding.
• Assuming stationarity for extended periods with dynamic stimuli can artificially inflate the apparent importance of correlations.

Conclusions:

• Neural decoding can be effectively simplified, with minimal information loss, even when ignoring complex correlations.
• The developed framework and metric I* provide a reliable method for assessing information in neural systems.
• Care must be taken when analyzing dynamic neural data to avoid misinterpreting the role of correlations due to inappropriate assumptions.