Dataset Distillation Technique Enabling ML On-board Training: Preliminary Results

Embedded systems with reduced computational capabilities suffer from the difficulties to re-train models on-board due to computational time and energy consumption strict constraints. Nevertheless, local training is of paramount importance to keep data local and private, avoid model drifts and to enable customization of models and federated learning. In this context, reducing the data size to be saved locally provide a twofold contribution: it limits the request in terms of memory and provide a much more compact representation for further train of the model, speeding up the whole process. In this work, we present the preliminary results of a dataset distillation technique that provides a data reduction up to 99%, enabling a local re-training of the ML model in a few numbers of epochs. It paves the way to the implementation of simplified algorithms of re-training on-board for microcontroller-based IoT devices, with an acceptable overall loss of accuracy.