Liquid Time Pocket
Liquid Time Pocket trains a recurrent cell whose hidden state relaxes toward a learned candidate through per-unit continuous time constants. The GRU and plain RNN controls contain exactly the same 1,887 parameters and receive delta time as an ordinary input feature.
All models train with gaps from 0.02 to 0.12, then face unseen gaps up to 0.40 on twice-longer sequences. The Space plots predictions against continuous timestamps.
Verified local result
At exactly 1,887 parameters, normal-gap RMSE was tightly matched: 0.0893 for the liquid cell, 0.0869 for the GRU, and 0.0881 for the RNN. Under unseen gaps up to 0.40, the liquid cell failed at 1.514 RMSE versus 0.274/0.281 for GRU/RNN. Its learned time constants ranged from 0.122 to 0.716, causing excessive phase-memory decay under the shifted interval distribution.
uv run python projects/liquid-time-pocket/train.py
uv run pytest tests/test_liquid_time_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.