Representation learning in the artificial and biological neural networks underlying sensorimotor integration

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Deep reinforcement learning (RL) offers a framework for understanding reward-based learning and decision-making.
• Investigating neural representations in both artificial agents and biological systems can reveal common principles of intelligence.

Purpose of the Study:

• To compare representation learning in deep RL agents and mice performing the same sensorimotor task.
• To explore the functional convergence between artificial and biological neural networks.

Main Methods:

• Trained deep RL agents and mice on a complex sensorimotor task with high-dimensional state and action spaces.
• Analyzed neural network models and posterior parietal cortex (PPC) neural activity.
• Utilized extensive hyperparameter search for model evaluation.

Main Results:

• Learning-dependent representations in optimized deep RL agents closely matched PPC neural activity.
• These representations were crucial for task performance in both agents and mice.
• PPC neurons showed representations of internally defined subgoals, similar to RL algorithms.

Conclusions:

• Deep RL models can capture key aspects of biological neural representations in sensorimotor tasks.
• The functional convergence highlights shared mechanisms in intelligent systems.
• This interdisciplinary approach provides new insights into neural computation and decision-making.