TinyRCE Forward Learning Under Tiny Constraints

Watch a technical symposium presentation exploring an innovative forward-only learning approach for ultra-low power embedded devices. Delve into the development of a hyper-spherical classifier designed specifically for microcontrollers and sensors, presented by Danilo PAU, Technical Director at STMicroelectronics. Learn about the classifier's architecture, which utilizes a convolutional neural network with randomly initialized weights or offline backpropagation training, requiring only 76.45 KiBytes of ROM and 40.2 KiBytes of RAM. Discover how the system processes real-time sensor data streams, implements a forget mechanism for redundant neurons, and achieves competitive performance against traditional supervised convolutional topologies. Examine practical applications through case studies in human activity monitoring and ball-bearing anomaly detection, with performance evaluations using PAMAP2, SHL, and CWRU datasets.