An interactive browser-based laboratory for exploring Liquid Neural Networks (LNNs) — the brain-inspired continuous-time adaptive architecture developed at MIT CSAIL by Ramin Hasani, Mathias Lechner, and Daniela Rus.
Inspired by the 302-neuron C. elegans nematode brain, liquid neural networks use differential equations with learnable time constants that adapt their dynamics based on input — enabling real-time adaptation, extreme parameter efficiency, and continuous-depth computation.
Explore a single Liquid Time-Constant neuron. Adjust parameters to see how the adaptive time constant τ(x, I) enables variable-timescale processing.
Compare continuous-depth (Neural ODE) vs discrete-layer computation. Visualize Euler vs RK4 trajectories through learned vector fields (spiral, saddle, limit cycle).
Interactive force-directed visualization of the 302-neuron nematode brain — the biological blueprint for liquid networks. Color-coded by neuron type.
The breakthrough that made liquid nets practical: analytical closed-form solutions eliminate expensive ODE solvers, achieving N× speedup with zero accuracy loss.
Watch a tiny liquid network adapt in real-time to concept drift (sine→square, slow→fast, clean→noisy) while a frozen LSTM fails. No retraining needed.
Compare Liquid vs LSTM vs Transformer vs Mamba on parameters, FLOPs, and memory. See why 19 neurons can drive a car.
- Single HTML file, vanilla JavaScript, HTML5 Canvas
- Zero dependencies, zero build step
- Deployed on Vercel
- Hasani et al. "Liquid Time-constant Networks" (AAAI 2021)
- Hasani et al. "Closed-form Continuous-time Neural Networks" (Nature Machine Intelligence 2022)
- White et al. "The Structure of the Nervous System of C. elegans" (1986)
- Liquid AI — https://liquid.ai
MIT