Analogue Neural Networks Achieve Low-Power Continuous Control with Trainable Connections.

Traditional physical neural networks, while promising for low-power machine learning, often constrain nonlinear device responses to act as simple scalar weights. This new research proposes an alternative architecture inspired by Kolmogorov-Arnold networks, where trainable nonlinear functions are placed directly on the connections between neurons. This design transforms each physical connection into a learnable computational element. These nonlinear functions are realized as analogue band-pass filters on field-programmable analogue arrays. The study found that this approach offers significant parameter efficiency for tasks involving smooth, continuously valued targets, such as robotic kinematics, continuous control, and photovoltaic maximum-power-point tracking. However, it provides no advantage for classification-like decision boundaries. The trained networks successfully transferred to hardware with high fidelity across approximately 35,000 connections. A dedicated CMOS implementation is projected to consume only about 30 microwatts, highlighting the potential for ultra-low-power AI. The findings suggest that the benefit stems from the placement of trainable nonlinearity on connections, rather than from a specific device technology.