28-m W Fully Embedded AI Techniques with On-site Learning for Low-Power Handy Tactile Sensing System

We have been studying an edge-AI learning-and-inference techniques for low-power handy battery operating tactile sensing system. It works in an event-driven of normally off power supply mechanism. The power consumption during classify operation at 10 inference per second is approximately 28m W. The proposed system adopts on-site embedded learning for the low power consumption and the high recognition accuracy, which is effective to operate in variety of environments. On-site learning function improves the recognition accuracy to classify the six kinds of paper from 50% to 91 % though humans can hardly identify them.