NSF grant $1.8M: decoding how the living brain rewrites only the right connections — toward AI that does not forget (UC San Diego)

This award is a prime example of the two-way street between AI and neuroscience: decode the rules by which the living brain rewrites its knowledge, and translate them into design principles for AI that does not forget.

1Three weaknesses of today's AI

The starting point is a weakness of current AI. Today systems excel at many tasks but need enormous data, struggle to keep learning new information without forgetting old (continual learning), and rely on rules far from the brains. That failure — catastrophic forgetting — is a deep problem for AI running in the real world. This project goes to neuroscience for a grounded answer.

2When and where rewiring is allowed

The central hypothesis is sharp: efficient learning depends on connectivity update rules that determine not only how neural activity modifies synaptic strength, but also when and where those modifications are permitted.

The brain does not blindly rewrite every connection; it selectively modifies only the appropriate ones while preserving existing knowledge — and the rules of that selection may be the key to preventing forgetting.

3Two-photon imaging and optogenetics

The methods are ambitious. An all-optical approach combining two-photon calcium imaging (observing many neurons at once), single-cell holographic optogenetic stimulation (activating a single targeted cell with light), and functional connectivity mapping both measures and experimentally induces plasticity in awake mice at single-cell resolution.

It resolves how behavioral state, neuromodulation, and local inhibitory circuits regulate these update rules, and how network organization and the functional identity of neuronal ensembles influence which connections are rewritten.