Schema formation in a neural population subspace underlies learning-to-learn in flexible sensorimotor problem-solving

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• Learning-to-learn is a fundamental cognitive process involving accelerated skill acquisition through experience with similar tasks.
• Understanding the neural mechanisms of learning-to-learn is crucial for both neuroscience and artificial intelligence research.
• The prefrontal cortex is implicated in complex cognitive functions, including learning and adaptation.

Purpose of the Study:

• To investigate the neural mechanisms underlying learning-to-learn dynamics.
• To model the emergence and reuse of neural schemas in a computational framework.
• To explore the role of plasticity in preserving and adapting learned representations.

Main Methods:

• Trained a recurrent neural network (RNN) model on a series of arbitrary sensorimotor mapping tasks.
• Analyzed the network's learning trajectory, observing an exponential speedup characteristic of learning-to-learn.
• Investigated the emergence of low-dimensional neural schemas within the RNN's population activity.
• Examined how the reuse of these schemas impacts learning efficiency and plasticity.

Main Results:

• The RNN model exhibited an exponential time course of accelerated learning, mimicking learning-to-learn.
• A neural schema, represented in a low-dimensional subspace of population activity, emerged during training.
• Schema reuse in new problems significantly facilitated learning by minimizing connection weight adjustments.
• Weight-driven modifications of the network's vector field were highlighted as critical for preserving and reusing schemas.

Conclusions:

• Recurrent neural network models can effectively capture learning-to-learn dynamics observed in biological systems.
• The emergence and reuse of low-dimensional neural schemas are key mechanisms for efficient learning and adaptation.
• Plasticity in neural networks plays a vital role in preserving learned knowledge while adapting to new challenges.
• This work provides insights into the computational principles of knowledge acquisition and generalization in the brain.