Neuroplasticity and Continual Learning in AI

Neuroplasticity—the brain's ability to reorganize and adapt throughout life—inspires approaches to continual learning that could solve AI's catastrophic forgetting problem.

Biological Plasticity

Types of Plasticity

Synaptic plasticity: Changes in connection strength
Structural plasticity: New synapses and dendrites
Neurogenesis: New neurons (in some regions)
Functional reorganization: Repurposing brain areas

Hebbian Learning

"Neurons that fire together, wire together":

Coincident activity strengthens connections
Local learning rule
Foundation for associative memory

Homeostatic Plasticity

Maintaining stable activity levels:

Scaling synaptic weights
Adjusting excitability
Preventing runaway potentiation/depression

Catastrophic Forgetting in AI

Current neural networks:

Forget old tasks when learning new ones
Overwrite previous weights
Lack stability-plasticity balance

Bio-Inspired Solutions

Elastic Weight Consolidation

Protecting important weights:

Estimate parameter importance
Penalize changes to important parameters
Inspired by synaptic consolidation

Progressive Neural Networks

Growing new capacity:

Freeze old networks
Add new columns for new tasks
Lateral connections transfer knowledge

Memory Replay

Rehearsing old experiences:

Sleep replays in hippocampus
Experience replay in RL
Generative replay

Sparse Representations

Minimizing interference:

Sparse coding in brain
Sparse activations in networks
Task-specific subnetworks

Future Directions

Always-learning AI systems
Models that improve with use
Efficient online adaptation
Biological-scale continual learning

Neither "Solution" Actually Solves It

This post lists Elastic Weight Consolidation and Progressive Neural Networks under "Bio-Inspired Solutions" to catastrophic forgetting. Both are real, useful, and neither has actually solved the problem — it's worth being specific about how each one fails, because the failure mode changes what a reader should expect from them.

EWC works by penalizing changes to parameters the network judged important for earlier tasks. Over a long sequence of many tasks, those penalties accumulate: each new task adds its own set of protected parameters, and eventually so much of the network is penalized against change that it can no longer learn new tasks at all. That's not catastrophic forgetting anymore — it's the mirror-image failure, catastrophic rigidity, and it's a direct, unavoidable consequence of how the method works, not an edge case.

Progressive Neural Networks avoid forgetting by never overwriting anything: each new task gets its own new column of the network, connected laterally to the columns before it. That genuinely eliminates forgetting, but only by making network size grow without bound as tasks accumulate — the "always-learning AI systems" this post's future directions section describes would need infinite capacity under this approach, which is not a workable path to the goal it's describing.

The honest state of the field: an approach that avoids both catastrophic forgetting and unbounded growth, at once, is still an open research problem, not something either bio-inspired method listed here has delivered. Each method solves one side of the trade-off by accepting the other, and a reader evaluating either for a real system needs to know which failure mode they're choosing, not just that biology "inspired" a fix.

References

The brain never stops learning—our AI should aspire to the same adaptability.