Skip to Content

Tutorial

T12 Learning Without Neurons

Traditionally, materials and machines are designed to exhibit desired responses to external signals.

Traditionally, materials and machines are designed to exhibit desired responses to external signals. However, a recent fusion of ideas from neuroscience and machine learning has given rise to a novel approach called physical learning. In this paradigm, systems, referred to as 'learning machines,' aren't explicitly designed for specific functions. Instead, they adapt to external signals, gradually developing the desired functionality through physical processes. Importantly, these systems adapt autonomously without any external modification of their elements.

Learning machines operate using physically realizable learning rules, that is, their elements evolve using local information and no explicit knowledge of the desired functionality. This innovative approach has been successfully demonstrated in laboratory experiments involving electronic circuits, elastic networks, and self-folding sheets. With physical learning, new classes of smart, programmable metamaterials can be created, capable of adapting to users' needs on the spot. Additionally, learning machines permit exploration of ideas from biology and computer science in a physical but tractable context. The principles and frameworks of physical learning are highly versatile and can be applied to a wide range of physical systems. This tutorial will feature speakers from the fields of soft condensed matter physics and biophysics. They will give an introduction to physical learning machines, their implementations, and their applications and connections with other topics, along with a discussion of the open questions and challenges in this emerging field.

Topics

  • Theory: Learning vs. design, local learning rules, contrastive learning methods, physical models of learning machines, physical implications of learning, physical regularization.
  • Experimental realizations: Electronic systems, Nonlinear learning, Spring networks, Self-folding sheets, Biological models (Physarum Polycephalum).
  • Applications: Neuromorphic computing, Smart metamaterials, Energy efficiency, Continual learning, Robust learning, Discovery of learned function.

Who should attend?

Graduate students, post-docs and other researchers in the fields of soft matter physics, biophysics and metamaterials interested in an introduction to the exciting new field of physical learning. The lectures will provide a pedagogical introduction to the general principles of physical learning and their application to various types of physical and biological systems in theory and experiment.

Organizers

  • Menachem Stern, AMOLF
  • Sam Dillavou, University of Pennsylvania

Presenters

  • Karen Alim, Technical University of Munich
  • Varda F. Hagh, University of Illinois, Urbana-Champaign
  • Sam Dillavou, University of Pennsylvania
  • Menachem Stern, AMOLF