Incremental On-Device Tiny Machine Learning

Tiny Machine Learning (TML) is a novel research area aiming at designing and developing Machine Learning (ML) techniques meant to be executed on Embedded Systems and Internet-of-Things (IoT) units. Such techniques, which take into account the constraints on computation, memory, and energy characterizing the hardware platform they operate on, exploit approximation and pruning mechanisms to reduce the computational load and the memory demand of Machine and Deep Learning (DL) algorithms. Despite the advancement of the research, TML solutions present in the literature assume that Embedded Systems and IoT units support only the inference of ML and DL algorithms, whereas their training is confined to more-powerful computing units (due to larger computational load and memory demand). This also prevents such pervasive devices from being able to learn in an incremental way directly from the field to improve the accuracy over time or to adapt to new working conditions. The aim of this paper is to address such an open challenge by introducing an incremental algorithm based on transfer learning and k-nearest neighbor to support the on-device learning (and not only the inference) of ML and DL solutions on embedded systems and IoT units. Moreover, the proposed solution is general and can be applied to different application scenarios. Experimental results on image/audio benchmarks and two off-The-shelf hardware platforms show the feasibility and effectiveness of the proposed solution