Machine learning at the mesoscale: a computation-dissipation bottleneck

TL;DR

This paper explores the trade-offs between performance and energy consumption in mesoscopic systems used as input-output devices, highlighting how non-equilibrium conditions can enhance performance through a computation-dissipation bottleneck.

Contribution

It introduces a framework analyzing the interplay of information processing, energetic costs, and irreversibility in mesoscopic systems, supported by real and synthetic data.

Findings

Non-equilibrium conditions improve system performance.

A fundamental trade-off exists between information compression and dissipation.

Non-reciprocal interactions induce dynamic irreversibility.

Abstract

The cost of information processing in physical systems calls for a trade-off between performance and energetic expenditure. Here we formulate and study a computation-dissipation bottleneck in mesoscopic systems used as input-output devices. Using both real datasets and synthetic tasks, we show how non-equilibrium leads to enhanced performance. Our framework sheds light on a crucial compromise between information compression, input-output computation and dynamic irreversibility induced by non-reciprocal interactions.