Exploring Unsupervised Learning on STM32 F4 Microcontroller

This paper investigated the application of unsupervised learning on a mainstream microcontroller, like the STM32 F4. We focused on the simple K-means technique, which achieved good accuracy levels on the four test datasets. These results are similar to those obtained by training a k-nearest neighbor (K-NN) classifier with the actual labels, apart from one case, in which K-NN performs consistently better. We propose an autonomous edge learning and inferencing pipeline, with a K-NN classifier which is periodically (i.e., when a given number of new samples have arrived) trained with the labels obtained from clustering the dataset via K-means. This system performs only slightly worse than pure K-means in terms of accuracy (particularly with small data subsets), while it achieves a reduction of about two orders of magnitude in latency times. To the best of our knowledge, this is the first proposal of this kind in literature for resource-limited edge devices.