NeST: Neuromorphic Stochastic Thermodynamics

Navigation Home Agenda Speakers Participants Presentations Background Papers

Organizers

David Wolpert, SFI, and Shantanu Chakrabartty, Washington University in St. Louis

Working Group Dates

December 10-12, 2025

Meeting Description

Neuromorphic Computing is a field that started three decades ago, inspired by the efficiencies of the information processing performed by biological brains. Its main goal is to replicate the complex architecture and functionality of biological neural circuits, but using in-silico circuits. There are several reasons for pursuing this goal. One is to create an alternative hardware platform for current computational tasks, especially those related to machine learning and AI. Over the past two decades, significant progress has been made in this area, with modern deep-learning and neuromorphic hardware already achieving, and even surpassing, human-level recognition performance in certain tasks. Another motivation comes from the thermodynamic inefficiency of present-day (Von Neumann style) computer architectures. In such architectures, a major energy cost—and source of heat—comes from transferring data between storage and processing units. In contrast, biological brains store data right where it is processed, thereby avoiding this problem. This suggests that replicating neurological architecture using in-silico circuits could lead to substantial energy savings over current computer architectures.

The field of stochastic thermodynamics is a recently developed body of work that extends conventional statistical physics to systems that are arbitrarily far from thermal equilibrium, with arbitrarily many degrees of freedom that are all changing on fast timescales. In addition, recent (and on-going) research in stochastic thermodynamics has produced a suite of unavoidable lower bounds on the energetic cost of any system that operates sufficiently quickly (``thermodynamic speed limit theorems"), sufficiently precisely (``thermodynamic uncertainty relations) or across a modular architecture (``mismatch cost). These lower bounds provide us with strengthened versions of the second law, whose strength grows larger the harder we try to design physical systems to meet any one of a wide variety of computational properties. These features of stochastic thermodynamics make it perfectly suited to analyze the thermodynamics and energy-scaling properties of both neuromorphic systems and the neurobiological systems that inspired them.

The objective of the working group is to bring together experts from the neuromorphic computing community and the stochastic thermodynamics community to facilitate convergence between the two disciplines.

Supported by a grant from the National Science Foundation award no: 2529902.