Onboard Class Incremental Learning for Resource-Constrained scenarios using Genetic Algorithm and TinyML

Deploying Machine Learning (ML) models in real-world settings over resource-constrained edge devices has always been a challenging task. While TinyML tackles this issue to an extent, by mostly using pre-trained Deep Learning (DL) models, the static nature of such models renders them ineffective for non-stationary data. A model having a low memory footprint that could allow onboard Class Incremental Learning (CIL) so as to accommodate data from new classes and also avoid catastrophic forgetting, is thus, the need of the day. The work described in this paper endeavours to meet this need by providing a method that utilises TinyML to accommodate a DL model onboard a resource-constrained device. To enable onboard CIL over the DL model, the method leverages Latent Replays (as exemplars) and a Genetic Algorithm (GA) to create a multi-fitness landscape that facilitates the inclusion of new class data, suppresses catastrophic forgetting and keeps a check on the quality of exemplars. Experiments and comparisons conducted for a gesture recognition task using time-series data with the proposed method deployed on a microcontroller, show the effectiveness of augmenting TinyML DL models with the GA for onboard CIL.