bptt-vs-rtrl-locality-tradeoff

IN premise — entries/2026/06/21/wiki-Recurrent_neural_network-chunk-4.md

Created 2026-06-21T09:55:53+00:00

BPTT is O(weights) per step but requires storing all activations (local in space, not time); RTRL is O(hidden × weights) per step but is online (local in time, not space).