Energy-aware Tiny Machine Learning for Sensor-based Hand-washing Recognition

Tiny wearable devices are nowadays one of the most popular and used devices in everyday life. At the same time, machine learning techniques have reached a level of maturity such that they can be used in the most varied fields. The union of these two technologies represents a valuable opportunity for the development of pervasive computing applications. On the other hand, pushing the machine learning inference on a wearable device can lead to nontrivial issues. In fact, devices with small size and low-energy availability, like those dedicated to wearable platforms, pose strict computational, memory, and power requirements which result in challenging issues to be addressed by designers. The main purpose of this study is to empirically explore the trade-off between energy consumption and classification accuracy of a machine learning-based hand-washing recognition task deployed on a real wearable device. Through extensive experimental results, obtained on a public human activity recognition dataset, we demonstrated that given an identical level of classification performance, a classic SVM classifier can save about 40% of energy with respect to a more complex LSTM network. Moreover, reducing the LSTM complexity, by lowering the number of its internal unit, can make the LSTM network energy cost-effective (with a savings of about 30%) at the cost of a reduction in accuracy of only 2%.