Online Learning Routines

Take a model from definition to an executable online-learning workflow. These chapters use compact tasks to make state handling, trace updates, and training behavior inspectable.

Note

Complete Quickstart and Core Concepts first if you have not yet compiled a BrainTrace model.

Choose a workflow

RNN Online Learning

Train a GRU on the copying task with D-RTRL, then compare the online workflow with a BPTT baseline.

Best for: continuous hidden states and sequence-memory tasks.

RNN Online Learning

SNN Online Learning

Build a recurrent LIF network, train it with pp-prop, and inspect how factorized traces differ from D-RTRL.

Best for: spike-based dynamics, surrogate gradients, and physical units.

SNN Online Learning

What both workflows establish

• how model state is initialized and reset between sequences;

• where braintrace.compile() enters the training pipeline;

• how repeated updates are executed with compiled stateful transforms; and

• which conclusions are specific to the demonstrated task and approximation.