Evolutionary data subset selection for class-Incremental learning on memory-constrained systems

Abstract

Training Machine Learning classifiers on extreme edge devices with non-volatile memory size ≤ 10 MB is challenging because of a) the small number of data examples that can be preserved on-device and b) the dynamic nature of the training dataset caused by the continual collection of new examples. Learning from a stream of data batches, each consisting of examples from a new class, is studied by class-incremental Continual Learning. We address the challenge of training an ML classifier on an extreme edge device that collects examples in a class-incremental fashion and propose a framework for selecting which examples should be preserved in non-volatile memory. We apply a genetic algorithm to identify data subsets with minimal degradation in the top-1 accuracy of a k-Nearest-Neighbors classifier. We evaluate the proposed subset selection method on MNIST and FashionMNIST datasets and observe top-1 accuracy gains of 26.3 % and 18.6 % relative to the mean accuracy obtained by random balanced subsets containing only 0.05 % of the training set. When the genetic algorithm was applied to a class-incremental setting, it achieved comparable accuracy to random balanced selection with 8× less non-volatile memory in specific cases.

More information

Identifiers

ORCID for Epifanios Baikas: orcid.org/0009-0007-0376-7636

ORCID for Danesh Tarapore: orcid.org/0000-0002-3226-6861

ORCID for David Thomas: orcid.org/0000-0002-9671-0917

Catalogue record

Export record

Altmetrics