Analog Neural Networks Hit 30 Microwatts by Making Each Connection a Learnable Filter

30 microwatts for a neural network that learns robotic kinematics and photovoltaic maximum-power-point tracking. That's the projection from a CMOS implementation of a new analog architecture that puts trainable nonlinearity on the connections, not just the nodes.

Why Physical Connections Need Their Own Nonlinearity

Most physical neural networks treat nonlinear device responses as scalar weights, forcing the analog physics into a rigid linear-combination-and-threshold model. This group instead borrows from Kolmogorov-Arnold networks: each physical connection becomes a learnable computational element. They realise these functions as analog band-pass filters on field-programmable analogue arrays (FPAAs).

The result is task-dependent, and the paper is honest about the boundary. On smooth, continuously valued targets - robotic kinematics, continuous control, solar MPPT tracking - the network needs far fewer nodes and connections than a multilayer perceptron. On classification-like decision boundaries, it offers zero parameter-efficiency advantage. The benefit comes from the smoothness of the physical basis matching the task.

35,000 Connections, Quantified Fidelity

Trained networks transfer to hardware across approximately 35,000 connections with quantified fidelity. That's not a simulation claim; they actually measured it. A memristive realisation reproduces the same behavior in simulation, confirming the advantage comes from placing trainable nonlinearity on connections, not from a specific device technology.

30 microwatts is the dedicated CMOS projection. That number changes the conversation around edge inference for continuous control - think drone motor mixing, solar panel angle adjustment, or robot joint mapping where a 100 mW microcontroller is overkill and a 10 W GPU is laughable. The tradeoff is clear: if your problem is smooth and continuous, this architecture eats MLPs for lunch on power and node count. If you need sharp decision boundaries, stick with digital.

The next step is building that CMOS chip and testing it on a real arm or panel. The 30 microwatt number is projected, not measured - but the FPAA transfer and memristive simulation suggest the physics holds up.

Source: Low-power analogue neural networks with trainable nonlinear connections for continuous control
Domain: arxiv.org