FAAST: Forward-Only Associative Learning via Closed-Form Fast Weights for Test-Time Supervised Adaptation

TL;DR

FAAST is a novel forward-only associative learning method that efficiently adapts pretrained models in constant time without memory overhead, matching or surpassing traditional methods in image and language tasks.

Contribution

It introduces FAAST, a new adaptation technique that analytically compiles labeled examples into fast weights in a single pass, eliminating the need for backpropagation or memory-based context.

Findings

Reduces adaptation time by over 90%

Saves memory usage by up to 95%

Matches or exceeds backprop-based adaptation performance

Abstract

Adapting pretrained models typically involves a trade-off between the high training costs of backpropagation and the heavy inference overhead of memory-based or in-context learning. We propose FAAST, a forward-only associative adaptation method that analytically compiles labeled examples into fast weights in a single pass. By eliminating memory or context dependence, FAAST achieves constant-time inference and decouples task adaptation from pretrained representation. Across image classification and language modeling benchmarks, FAAST matches or exceeds backprop-based adaptation while reducing adaptation time by over 90% and is competitive to memory/context-based adaptation while saving memory usage by up to 95%. These results demonstrate FAAST as a highly efficient, scalable solution for supervised task adaptation, particularly for resource-constrained models. We release the code and models at https://github.com/baoguangsheng/faast.