Event-Triggered Liquid Neural Networks for Computationally Efficient Time-Series Forecasting
Maya Assaf
Published: 2026/11/14 · PUB-2026-000052
Abstract
Liquid time-constant (LTC) networks adapt their dynamics to the input through ordinary differential equations, which suits irregularly sampled series, but their solver runs at every step even when the input hardly changes. We propose ET-LTC, an LTC cell whose solver is invoked only when the root-mean-square change of the standardised input since the last trigger exceeds a threshold; skipped steps use the exact solution of the LTC equation with the nonlinearity frozen, and fired steps use a step-doubling adaptive integrator. We derive a conditional state-error bound for skipped intervals and check it numerically on random cells using synthetic inputs; this check is separate from the dataset experiments. Across five dataset variants (Jena Climate, UCI Electricity, each with 50% simulated irregularity, and PhysioNet 2012) and ten seeds, ET-LTC used 13% to 31% of the FLOPs of a six-unfold LTC with a trigger rate of 32% to 66%, at a test error 0.9% to 1.8% higher (RMSE) or 0.010 lower (AUROC). Non-inferiority within 2% held after Holm correction on four of five variants. A random gate with the same firing rate was clearly worse, and holding the state instead of the closed-form skip cost accuracy. However, an LTC with a single unfold matched ET-LTC at lower or similar cost on the irregular and clinical data. The method therefore helps mainly on smooth, regularly sampled signals.
Keywords: liquid time-constant networks، event-triggered computation، closed-form skip rule، adaptive ODE solver، irregularly sampled time series، time-series forecasting، computational efficiency، edge deployment، non-inferiority testing