Context selectivity with dynamic availability enables lifelong continual learning.

Area of Science:

• Neuroscience
• Machine Learning
• Cognitive Science

Background:

• The human brain exhibits remarkable lifelong learning (or continual learning, CL) capabilities, retaining skills over extended periods despite new learning experiences.
• The underlying neural mechanisms enabling this persistent memory and skill retrieval are not yet fully understood.

Purpose of the Study:

• To propose a novel, bio-plausible meta-plasticity rule that explains the brain's lifelong learning (CL) capacity.
• To formalize this rule in a neuro-centric model and evaluate its performance in simulation.

Main Methods:

• Development of a meta-plasticity rule based on two principles: context-selective neurons and a local availability variable that modulates plasticity.
• Neuro-centric formalization of these principles to create a computational model for CL.
• Simulation of the model on image recognition and natural language processing benchmarks for CL.

Main Results:

• The proposed model successfully balances forgetting and consolidation, crucial for effective lifelong learning (CL).
• The model demonstrated superior transfer learning performance compared to contemporary CL algorithms on benchmark tasks.
• Neuron selectivity and neuron-wide consolidation emerged as a viable hypothesis for enabling CL.

Conclusions:

• The proposed meta-plasticity rule offers a simple yet effective mechanism for enabling lifelong learning (CL) in neural systems.
• This neuro-centric approach provides a promising direction for understanding and replicating the brain's ability to learn continuously.
• The findings suggest potential applications in artificial intelligence for developing more robust and adaptable learning systems.