Clockwork RNN Pocket
A modern, parameter-matched retest of periodic recurrent computation inspired by the Clockwork RNN. Four eight-unit hidden blocks update at periods 1, 2, 4, and 8. The model competes with an ordinary tanh RNN and a smaller GRU on noisy three-timescale forecasting.
All variants train on length-64 sequences and are evaluated both at that length and zero-shot at length 256. The interactive Space overlays their predictions on the same generated signal.
Verified local result
All models contain exactly 1,153 parameters. At zero-shot length 256, Clockwork RNN reached 0.0570 RMSE, versus 0.0492 for the ordinary RNN and 0.0466 for the GRU. Periodic hidden updates did not improve this benchmark; the matched controls make that negative result explicit.
uv run python projects/clockwork-rnn-pocket/train.py
uv run pytest tests/test_clockwork_rnn_pocket.py
Hosted showcase
This free static Space preserves the complete original Gradio source, trained artifacts, evaluation files, and local launch requirements. Hugging Face now requires PRO for CPU-backed Gradio hosting, so the public landing page is static while the checked-in app.py remains the authoritative runnable demo.