Tiny Machine Learning For Printed Electronics

A new approach optimizes tiny machine learning for ultra-low-power, flexible printed electronics, enabling efficient neural networks for wearables, implants, and other applications.

Researchers from the Karlsruhe Institute of Technology and the University of Patras have developed a method to apply tiny machine learning (tinyML) to compact printed electronics, introducing “sequential super-tinyML multi-layer perceptron circuits” for multi-sensory applications.

Super-tinyML is designed to optimize machine learning models for ultra-low-power applications, such as wearable technologies and implants. These applications require flexibility, conformability, and non-toxicity—qualities that traditional silicon-based systems cannot provide. Printed Electronics (PE) meets these needs while enabling cost-effective, on-demand fabrication. However, Neural Networks (NN), with numerous features that are often essential for these applications, have been impractical in PE due to limitations such as large feature sizes and restricted device counts.

– Advertisement –

The team’s solution is a super-tinyML architecture aimed at application-specific neural networks, which they call sequential super-tinyMLs. This architecture consists of controller logic, a hidden layer, an output layer, and an argmax. The hidden and output layers contain the neurons of a Multi-Layer Perceptron (MLP), with some neurons using multi-cycle operations and others using single-cycle operations.

By employing a multi-cycle approach and neuron approximation, the team claims it is possible to reduce the resource demands of neural networks enough to implement effective models on printed electronics for wearables, implantables, and more. Their experiments suggest this approach achieves a nearly 36-fold increase in feature size and over 65 times the number of coefficients compared to competing methods. Overall, the team’s architecture boasts a 12.7× improvement in area efficiency and an 8.3× reduction in power consumption compared to the current state-of-the-art.