Introducing OscNet: A JAX library for oscillatory neural networks and dynamical systems.
OscNet provides a framework for building and training neural networks based on oscillatory dynamics — coupled oscillator networks, continuous-time neural networks, and general dynamical systems. Built on JAX and Equinox for differentiable, high-performance computation. https://github.com/samim23/oscnet
Features
- Oscillator models: Harmonic, Van der Pol, Stuart-Landau, Kuramoto, FitzHugh-Nagumo
- Coupling topologies: Hierarchical fractal, power-law, log-periodic
- Analysis tools: Edge-of-chaos, Floquet analysis, bifurcation, stability
- Training utilities: Criticality initialization, stochastic forcing, schedulers
- Visualization: Phase space, network dynamics, oscillator analysis
More on the projects background: In Resonance with Nature - Toward a New Kind of Machine Learning with Oscillatory Neural Networks