Liquid Time-constant Networks

Summary

LTCs replace a Neural-ODE f with linear first-order units whose τ is gated by f: dx/dt = −[1/τ+f]x + f A, solved with a fused implicit/explicit Euler step and trained by BPTT. τ and hidden state are bounded; trajectory-length expressivity exceeds Neural ODEs and CT-RNNs. AAAI-21. Improves several UCI-style series vs LSTM/CT-RNN/Neural ODE/CT-GRU (traffic SE 0.099 vs LSTM 0.169; person-activity setting 2 acc 0.882 vs Latent ODE 0.846). Code https://github.com/raminmh/liquid_time_constant_networks.

Keywords

ltc · neural-ode · continuous-time · time-series · aaai · mit · tu-wien · liquid-networks

Topics

neural ODEs, continuous-time RNNs, time-series

Research notes

• Primary: arxiv abs (cs.LG; also cs.NE, stat.ML). License CC BY 4.0 on HTML at check. Comment: Accepted to AAAI-21. Equal contrib Hasani/Lechner; Amini/Rus MIT CSAIL; Grosu TU Wien. Code on abs https://github.com/raminmh/liquid_time_constant_networks (1,855 stars at check; no LICENSE file at check). HF githubRepo auto-link Ipsedo/LiquidNetworks (29 stars) is not the abs repo. HF paper page 3 upvotes; unofficial linked models krystv/LiquidFlow and krystv/liquid-diffusion not copied into hf_* fields. Discord posted PDF. Uses public UCI/HAR/gesture/traffic/MuJoCo series rather than a new hosted corpus, so no datasets_local row.