raminmh/liquid-s4

Combines linearized liquid neural networks with state-space models to achieve efficient long-range sequence modeling across diverse domains (vision, audio, text, physiological signals). Implements two kernel variants—polynomial basis (PB) and Kronecker basis (KB)—with configurable polynomial degrees for trading off expressiveness and computational cost. Built on PyTorch-Lightning and Hydra for modular training, integrating custom Cauchy kernels or PyKeOps for accelerated kernel operations.

388 stars. No commits in the last 6 months.