Low-power analogue nets with KAN-like connections achieve 30µW control

A team led by Vidamour et al. has introduced a new approach to physical neural networks that promises ultra-low-power computation for continuous control tasks. Traditional analogue networks force nonlinear device responses to act as simple scalar weights, limiting expressiveness. Inspired by Kolmogorov-Arnold networks (KANs), the researchers instead placed trainable nonlinear functions directly on the connections—implementing these as analogue band-pass filters on field-programmable analogue arrays. This design turns each physical connection into a learnable computational element, allowing the network to represent smooth, continuously valued targets like robotic joint kinematics, continuous control policies, and solar maximum-power-point tracking with dramatically fewer nodes and connections than multilayer perceptrons.

Importantly, the benefit is task-dependent: the approach excels on smooth regression tasks but offers no parameter-efficiency advantage on classification-like decision boundaries. The team successfully transferred trained networks to hardware across approximately 35,000 connections with quantified fidelity. A dedicated CMOS implementation is projected to consume only about 30 microwatts, orders of magnitude less than conventional digital chips. A memristive realization reproduced the same behavior in simulation, confirming that the advantage comes from placing trainable nonlinearity on connections rather than a specific device technology. This work paves the way for extremely low-power edge AI capable of real-time control in robotics, autonomous systems, and energy management.