CAREER: Learning and Information Processing in Non-linear Mechanical Systems

NONTECHNICAL SUMMARY

This CAREER award will support research to understand how mechanical systems can learn intelligent behavior in ways similar to neural networks. Mechanical materials will be studied that actively re-organize themselves through local processes so as to recognize and respond to subtle patterns in physical stimuli. The fields of condensed matter physics and materials science have traditionally focused on systems with fixed capabilities.

Inspired by the brain, this project seeks to understand the principles that allow physical systems to change themselves and `learn’ new functions. The brain is remarkable not only in its ability to learn but in that learning is done without any global coordinator who decides which neurons should connect to whom in order to recognize a friendly face.

Instead, neurons wire up with each other through local processes, such as Hebbian learning (‘fire together, wire together’), and self-organize to recognize patterns in the world around us.

Inspired by such local self-organization in service of global function, the PI recently proposed training a material in using Hebbian-like learning to reinforce right behaviors and anti-Hebbian learning to penalize wrong behaviors. The ability to eliminate undesired behaviors in addition to enhancing desired ones will allow for a powerful physical analog of supervised learning in neural networks.

However, a material cannot autonomously learn in this Hebbian and anti-Hebbian manner without special forms of memory. Further, fundamental principles of thermodynamics dictate that these forms of memory and their erasure require materials to consume energy.

This project will investigate such non-equilibrium requirements for Hebbian-inspired learning in materials. The consideration of non-equilibrium phenomena is at the frontier of condensed matter physics, but instead of order and pattern formation, this project focuses on non-equilibrium information processing. Besides developing the underlying theory, this project will investigate how feedback loops between mechanics and chemistry can lead to non-equilibrium learning behaviors in practice.

The PI will also investigate what physical aspects make a mechanical system act like a neural network, i.e., be capable of recognizing complex patterns in stimuli.

This work will provide insight into how biological systems adapt to their environment since feedback between mechanics and chemistry is common in cell biology. At a fundamental level, this project will expand our view of what non-equilibrium mechanical systems can do beyond mechanical responses to learning behaviors like neural networks.

Undergraduate and graduate students will be trained in interdisciplinary research. The PI will develop courses that target the emerging scientific interface between physics and materials science. For the broader public, the PI will develop hands-on demos that illustrate and relate to cutting-edge ideas in both physics and theoretical computer science.

These demos will include 3-D printed networks and mechanical structures that change and adapt as they are used. They will communicate to the broader public how exciting new research emerges at the interface of different sciences. The demos will be disseminated through outreach events and to K-12 educators for use in the classroom.

TECHNICAL SUMMARY

The goal of this proposal is to elucidate the non-equilibrium requirements for mechanical systems to physically learn new functionalities in a way rivaling neural networks. While many existing frameworks for adaptive materials can enhance a desired behavior by lowering its energy during a period of training, these frameworks are missing key ingredients that we will address here.

Unlike artificial neural networks that are trained by global optimization, biological neural networks are thought to learn through local rules such as Hebbian learning. Since physical systems are also typically constrained by locality, the PI has explored training a material in a supervised way by using Hebbian learning when it shows the right behaviors and anti-Hebbian learning when it shows the wrong behaviors.

The ability to eliminate undesired behaviors in addition to enhancing desired ones will allow for a powerful physical analog of supervised learning in neural networks. However, for a material to learn autonomously through such Hebbian and anti-Hebbian learning, it must have forms of memory that are only allowed when detailed balance is broken. This project will determine the non-equilibrium memory requirements for supervised learning in matter that experiences desired and undesired behaviors over time.

In addition to abstract theory, this project will also explore how natural chemo-mechanical feedback circuits, found in numerous biological and synthetic systems, can implement such memory. Through collaborations with materials scientists, this project will instantiate the theory here in specific systems.

This work will expand our understanding of what non-linear non-equilibrium disordered mechanical systems can do. While non-equilibrium driving has been extensively studied as a mechanism of order, organization and pattern formation, our framework shows how non-equilibrium driving enables a system to learn statistical stimuli-response behaviors, much like a neural network.

A frontier in soft condensed matter is to reveal the universe of behaviors enabled by chemo-mechanical coupling; but developments are currently siloed with sophisticated neural network-like information processing investigated in well-mixed molecular circuits without any mechanics. This work will build a unified theoretical framework, combining statistical physics models of molecular circuits with adaptive mechanical networks, allowing us to compute the dissipation requirements for adaptive behaviors in mechanical systems.

Finally, by contrasting learning in mechanical systems and in neural networks, this work will provide a new perspective on learning itself.

The proposed theoretical condensed matter work will develop tools that will impact materials engineering, including metamaterials and bio-materials, and biological physics. While current approaches optimize for desired behaviors chosen ahead of time, the new framework here allows materials to autonomously learn and improve function over time by harnessing recent advances ranging from DNA nanotechnology to ceramics and alloys.

The interdisciplinary research here is also an ideal platform for impacting high school education through engaging demos based on 3-d printed meta-materials and shape-memory alloys. These demos will be disseminated through public science outreach events in Chicago.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.