TinyD\'ej\`aVu: Smaller RAM and Faster Inference with Neural Networks on MCUs for Sensor Data Streams
A new framework called TinyDéjàVu has been developed to significantly reduce the RAM requirements for neural network inference on microcontrollers. This framework can decrease RAM usage by up to 90% while maintaining similar compute latency compared to previous methods, making it highly efficient for battery-powered sensor devices. The implementation is open-source and has been benchmarked on common microcontroller hardware. AI
IMPACT Enables more complex neural network models to run on resource-constrained embedded systems, potentially expanding the capabilities of IoT devices.