Rate-Based RNN Examples#
These examples focus on rate-based recurrent models and comparisons with BPTT. The runnable scripts live in the repository’s examples directory.
100-gru-on-copying-task.py trains a GRU on the copying task and compares online diagonal RTRL with BPTT. API: GRUCell.
101-integrator-rnn.py trains a
MiniGRUintegrator and plots both predictions and training losses. API: MiniGRU.
The comparisons demonstrate behavior on the stated model and task. They do
not establish general gradient equivalence, because D_RTRL retains a
diagonal hidden-Jacobian approximation.
Begin with the RNN Online Learning workflow for the complete training sequence.