Orthogonal representations for robust context-dependent task performance in brains and neural networks.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• Neural populations must efficiently process multiple tasks, often with conflicting demands.
• Context-dependent decision-making requires flexible neural representations.
• Existing models propose 'lazy' and 'rich' coding strategies with different learning trade-offs.

Purpose of the Study:

• To investigate how neural populations code for multiple, potentially conflicting tasks.
• To define computational models of 'lazy' and 'rich' neural coding.
• To identify which coding strategy is implemented in biological brains.

Main Methods:

• Computational simulations using neural networks to model 'lazy' and 'rich' learning.
• Behavioral testing and neuroimaging (fMRI) in humans.
• Analysis of neural signals from macaque prefrontal cortex.

Main Results:

• Defined 'lazy' coding via random input expansion and 'rich' coding via structured representations.
• Identified low-dimensional, orthogonal manifolds as a key 'rich' coding solution.
• Observed neural coding patterns in humans and macaques consistent with the 'rich' learning regime.

Conclusions:

• Biological brains appear to employ a 'rich' coding strategy for context-dependent decision-making.
• This 'rich' coding involves structured neural representations and low-dimensional manifolds.
• Findings provide insights into the neural basis of cognitive flexibility and task management.