Towards implementing neural networks on edge IoT devices

Combining AI and IoT in small devices is challenging. Tokyo University of Science researchers developed a new training algorithm for binarized neural networks (BNNs) implemented in a computing-in-memory architecture for IoT devices. Their approach uses ternary gradients, an enhanced Straight Through Estimator, and probabilistic parameter updating. The design achieved over 88% accuracy on handwriting recognition, matching regular BNNs with faster convergence. This could enable more powerful and efficient IoT devices with improved AI capabilities for health monitoring and smart homes.