nested learning

A framework from Google Research that describes a model as a stack of learning problems running inside or alongside one another, each optimizing over its own stream of information at its own update rate, which puts network architecture and training algorithm on the same footing as optimization at different levels. Under this lens backpropagationbackpropagation, n.: The method neural networks use to learn from mistakes: it works backward from the error to figure out how much each internal connection contributed, so every connection can be adjusted in the right direction. and attentionattention, n.: A way for an AI model to choose what to focus on. For each word it processes, the model scores how relevant every other word is and draws information mostly from the relevant ones. both work as associative memories, and architectures differ chiefly in how they compress the information streaming through them. Arranging components by how fast they update gives a 'continuum memory system', a spectrum of modules from fast-changing to slow-changing, meant to let a model keep learning (continual learningcontinual learning, n.: The ability of an AI to keep learning new things over time from a stream of experience, without forgetting what it already knew and without being retrained from scratch.) with less catastrophic forgettingcatastrophic forgetting, n.: When a neural network trained on something new gets suddenly and badly worse at what it could already do, because the new training rewrites the internal settings the old skills relied on.. A proof-of-concept self-modifying recurrent architecture, Hope — a variant of Google's earlier Titans memory model — reported better language-modeling, reasoning and long-context retrieval results than comparable transformers and recurrent models.

Nested learning recasts the optimizer itself as just another memory with a slower clock.