Show HN: Neural Particle Automata
4 hours ago
- #neural cellular automata
- #particle systems
- #machine learning
- Neural Particle Automata (NPA) extends Neural Cellular Automata (NCA) to dynamic particle systems, with particles having continuous positions and internal states updated by a learnable neural rule.
- NPA uses differentiable Smoothed Particle Hydrodynamics (SPH) operators for local perception, enabling scalable end-to-end training and addressing challenges of dynamic neighborhoods and quadratic scaling.
- SPH perception replaces grid-based methods with smooth kernels to aggregate nearby particles, estimating quantities like density, gradients, and moment matrices to form a local perception vector.
- NPA demonstrates key NCA behaviors such as robustness and regeneration, while enabling new particle-specific behaviors in tasks like morphogenesis, point-cloud classification, and texture synthesis.