Bayesian continual learning and forgetting in neural networks.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Artificial neural networks (ANNs) struggle with catastrophic forgetting and remembering, unlike biological synapses.
• Existing methods lack a principled way to balance memory retention and flexibility.

Purpose of the Study:

• To introduce Metaplasticity from Synaptic Uncertainty (MESU), a Bayesian update rule for continuous learning in ANNs.
• To enable ANNs to combine learning and forgetting without explicit task boundaries, inspired by biological synapses.

Main Methods:

• Developed MESU, a Bayesian update rule scaling parameter learning by uncertainty.
• Incorporated epistemic uncertainty estimation for out-of-distribution detection.
• Utilized weight sampling for predictive statistics computation.

Main Results:

• MESU mitigates forgetting while preserving plasticity in image-classification benchmarks.
• Outperformed established synaptic-consolidation methods on sequential Permuted-MNIST tasks.
• Demonstrated superior performance over conventional techniques in task-incremental CIFAR-100.

Conclusions:

• MESU offers a biologically inspired approach to robust, perpetual learning in ANNs.
• Connects metaplasticity, Bayesian inference, and Hessian-based regularization.
• Provides a pathway for ANNs to achieve continuous learning akin to biological systems.