Show HN: Neural Particle Automata (selforg-npa.github.io)

A groundbreaking development in the AI community has emerged with the introduction of Neural Particle Automata (NPA), a new model that transitions from traditional Neural Cellular Automata (NCA) to dynamic particle systems. Unlike NCAs, which operate on static grids, NPAs treat each cell as a particle with a continuous position and internal state heavily influenced by a shared learnable neural rule. This innovation enhances cell individuation and enables heterogeneous dynamics, thereby prioritizing computational resources where activity occurs. However, the dynamic nature of particle neighborhoods presents challenges, particularly in scaling local interactions efficiently. The team has tackled these issues by implementing differentiable Smoothed Particle Hydrodynamics (SPH) for neighborhood perception, paired with memory-efficient CUDA kernels for improved training scalability. The significance of NPA lies in its ability to maintain the robustness and regenerative power of NCAs while introducing capabilities tailored to particle systems. By allowing particles to exist within irregular configurations, NPAs showcase potential applications across various tasks, including morphogenesis, point-cloud classification, and particle-based texture synthesis. The approach leverages SPH to create a compact local perception vector that retains locality while adapting to the fluid dynamics of particles. These advancements position NPA as a promising neural model for self-organizing systems, paving the way for innovative developments in AI and machine learning.