liquid time-constant networks

**Liquid Time-Constant Networks (LTCs)** are a **class of continuous-time Recurrent Neural Networks (RNNs)** — created by Ramin Hasani et al., where the hidden state's decay rate (time constant) is not fixed but varies adaptively based on the input, inspired by C. elegans biology. **What Is an LTC?** - **Definition**: Neural ODEs where the time-constant $ au$ is a function of the input $I(t)$. - **Equation**: $dx/dt = -(x/ au(x, I)) + S(x, I)$. - **Behavior**: The system can be "fast" (react quickly) or "slow" (remember long term) dynamically. **Why LTCs Matter** - **Causality**: They explicitly model cause-and-effect dynamics governed by differential equations. - **Robustness**: Showed superior performance in driving tasks, generalizing to uneven terrain better than standard CNN-RNNs. - **Interpretability**: Sparse LTCs can be pruned down to very few neurons (19 cells) that are human-readable (Neural Circuit Policies). **Liquid Time-Constant Networks** are **adaptive dynamical systems** — robust, expressive models that bridge the gap between deep learning and control theory.