AgileNN is the first work that achieves real-time inference (<20ms) of mainstream neural network models (e.g., ImageNet) on extremely weak MCUs (e.g., STM32 series with <1MB of memory), without impairing the inference accuracy. The usage of eXplainable AI (XAI) techniques allows >6x improvement of feature compressibility during offloading and >8x reduction of the local device’s resource consumption.
