Low-power analogue neural networks with trainable nonlinear connections for continuous control

2026-06-21 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Hardware Architecture, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

A new approach to physical neural networks (PNNs) introduces trainable nonlinear functions on connections, drawing inspiration from Kolmogorov-Arnold networks. Unlike most PNN architectures that use nonlinear device responses as scalar weights, this method implements these functions as analogue band-pass filters on field-programmable analogue arrays. This design significantly enhances parameter-efficiency for tasks requiring smooth, continuously valued targets, such as robotic kinematics, continuous control, and photovoltaic maximum-power-point tracking. These networks achieve comparable performance with far fewer nodes and connections than traditional multilayer perceptrons. While effective for continuous tasks, they offer no parameter-efficiency advantage for classification-like decision boundaries. The trained networks demonstrate high-fidelity transfer to hardware across approximately 35,000 connections, with a dedicated CMOS implementation projected to consume only 30 microwatts. Simulations confirm the advantage stems from placing trainable nonlinearity on connections, independent of the specific analogue device.

Key takeaway

For robotics engineers developing low-power continuous control systems, you should investigate analogue neural networks with trainable nonlinear connections. This architecture significantly reduces node and connection counts compared to MLPs for smooth, continuously valued tasks like kinematics. Consider prototyping with field-programmable analogue arrays to leverage the projected 30 microwatt operation, especially where power efficiency is critical. Your designs could achieve high fidelity with fewer physical components.

Key insights

Trainable nonlinear connections in analogue neural networks offer high parameter-efficiency for continuous control tasks.

Principles

• Smooth physical basis benefits continuous targets.
• Trainable nonlinearity on connections is key.
• Task-dependent efficiency for PNN architectures.

Method

Realize trainable nonlinear functions as analogue band-pass filters on field-programmable analogue arrays.

In practice

• Apply to robotic kinematics.
• Use for continuous control systems.
• Optimize photovoltaic power tracking.

Topics

• Analogue Neural Networks
• Trainable Nonlinearity
• Continuous Control
• Field-Programmable Analogue Arrays
• Low-Power AI Hardware
• Robotic Kinematics

Best for: Research Scientist, AI Scientist, AI Hardware Engineer, Robotics Engineer

Related on AIssential

• Energy-Aware Spike Budgeting for Continual Learning in Spiking Neural Networks for Neuromorphic Vision — An energy-aware spike budgeting framework improves SNN continual learning by adaptively balancing accuracy and energy across vision modalities.
• Modern analog computing for solving differential and matrix equations — Modern analog computing, utilizing CMOS and resistive memory, offers efficient solutions for differential and matrix equations in data-intensive applications.
• Optical Implementation of Equilibrium Propagation Using Spatial Photonic Ising Machines — A Spatial Photonic Ising Machine enables energy-efficient optical-digital Equilibrium Propagation for training energy-based networks.
• An affordable hardware-aware neural architecture search for deploying convolutional neural networks on ultra-low-power computing platforms — A lightweight hardware-aware NAS generates tiny, accurate CNNs for ultra-low-power microcontrollers, enabling efficient embedded AI.
• A Synthesizable RTL Implementation of Predictive Coding Networks — Predictive coding offers a local, hardware-friendly alternative to backpropagation for distributed online learning.
• SpikeDecoder: Realizing the GPT Architecture with Spiking Neural Networks — SpikeDecoder implements energy-efficient Transformer decoders using SNNs for NLP, significantly reducing power consumption.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.