Ehsan Pajouheshgar
Dr. Pajouheshgar is a postdoctoral researcher at EPFL Lausanne, working jointly with the Image and Visual Representation Laboratory (IVRL) and the Chair of Statistical Field Theory (CSFT).
Abstract — “Neural Cellular Automata (NCA) offer a simple but powerful idea: encode local update rules as neural networks and learn them with gradient descent to achieve a desired global behavior. In this talk, I will argue that this combination of locality, differentiability, and learning turns NCAs into a practical framework for discovering self-organising systems that are hard to design by hand or even anticipate in advance.
I will present a series of examples where we use this framework to learn dynamical rules that give rise to complex textures, robust memories, complex morphologies, and particle-based collective behaviours, all from high-level objectives rather than hand-crafted mechanisms. Across these domains, a common picture emerges: by constraining only locality and shared parameters while optimising with gradient descent, NCAs reliably uncover rich, robust, and sometimes surprising self-organising dynamics, suggesting a general recipe for exploring the space of possible dynamical laws.”
We highly recommend playing around with the various beautiful and intuitive web demos Dr. Pajouheshgar created to illustrate his work: DyNCA, NoiseNCA, Cells to Pixels, MemoryNCA, and MeshNCA.