Edge-AI for Power reduction: Application to accelerometer sensors

International audienceEdge-AI is the use of AI algorithms directly embedded on a device contrary to a remote AI which makes use of an AI on a cloud or remote server for prediction. Recent improvements in microcontroller computing capabilities along with deep learning algorithms conversion frameworks made it easier to run small models directly on microcontroller units. In this paper, we present how an embedded deep convolutional neural network can be used for real-time human activity recognition with +98% accuracy and extending battery life. Experiments conducted on an Arm Cortex-M4 showed that average power can be reduced up to 10% when inferences are run on edge vs using a remote server