Liquid Time-constant Networks

Introduced liquid neural networks — continuous-time recurrent networks with time-constants that vary as a function of the hidden state and input. The dynamics adapt to the input, enabling compact, interpretable models for time-series tasks.

The foundational paper for Liquid AI's architecture. Demonstrated stronger performance on time-series forecasting with far fewer parameters than LSTMs and standard RNNs. AAAI 2021. By Hasani, Lechner, Amini, Rus, and Grosu.