Abstract
Adaptivity and learning are ubiquitous in living systems, from cytoplasm and biofilms to animal flocks. However, the problem of designing nonliving mechanical systems capable of self-learning behavior continues to present theoretical and experimental challenges. Here we introduce self-learning mechanical circuits: passive systems which take stochastic input forces from changing environments and constantly update their internal state in response, thus representing a mechanical unsupervised learning unit. Motivated by simplified models of synapses, our circuits are composed of adaptive directed springs, which change stiffness in a directional manner. We provide an experimental realization of these elastic learning units and demonstrate how stability arises from adaptivity in certain network architectures. In particular, we theoretically demonstrate the ability of a network to autonomously uncover patterns in external inputs. Our results have applications towards the design of adaptive materials that can self-optimize in different environments.
Funding Information
  • Schmidt Sciences
  • Gordon and Betty Moore Foundation
  • Dalio Foundation
  • Stanford Science Fellowship

This publication has 68 references indexed in Scilit: