forward-only

FAAST proposes a forward-only method that compiles labeled examples into fast weights analytically, enabling efficient test-time supervised adaptation without backpropagation, achieving over 90% speedup and 95% memory savings while maintaining performance.