Signatures of task learning in neural representations

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Neural plasticity is the basis of learning.
• Large-scale neural recording techniques enable studying coordinated neural activity.
• Artificial neural networks offer insights into multi-task and continual learning.

Purpose of the Study:

• To review recent findings on task learning at the population level.
• To propose a new framework for understanding task learning.
• To integrate insights from biological and artificial neural circuits.

Main Methods:

• Review of recent findings in biological and artificial circuits.
• Analysis of neural plasticity and learning-induced activity changes.
• Examination of neural manifold geometry and latent dynamics.

Main Results:

• Learning involves coordinated activity changes across neurons.
• Task learning can be understood through evolving neural manifold geometry.
• A tradeoff between non-interference and compositionality guides flexible multi-task learning.

Conclusions:

• A new framework integrating biological and artificial findings advances population-level task learning understanding.
• Coordinated neural activity changes and manifold geometry are key to learning.
• Principles from artificial intelligence inform understanding of biological neural circuit flexibility.