Local prediction-learning in high-dimensional spaces enables neural networks to plan

Area of Science:

• Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Planning and problem-solving are crucial higher brain functions, yet the underlying neural mechanisms remain largely unknown.
• Current AI planning methods often require extensive data and computational resources.

Purpose of the Study:

• To elucidate the neural basis of planning and problem-solving.
• To propose a novel, efficient learning mechanism for cognitive map acquisition.

Main Methods:

• A cognitive map learning framework based on predicting the next observation via local synaptic plasticity.
• Developing a system that learns relations between actions and observations.

Main Results:

• The learned cognitive map provides a quasi-Euclidean sense of direction, enabling efficient online planning comparable to AI algorithms.
• The method automatically extracts environmental regularities for generalization in physical spaces, accelerating learning in navigation and locomotion tasks.
• The proposed learner functions without a teacher, backpropagation, or large datasets, akin to self-attention networks.

Conclusions:

• Learning a predictive cognitive map is sufficient for complex planning and problem-solving.
• This approach offers a blueprint for energy-efficient neuromorphic hardware capable of autonomous learning and advanced cognitive functions.