Andrii Krutsylo
PhD student at the Institute of Computer Science of the Polish Academy of Sciences
The study aims to reduce the decline in performance of a model trained incrementally on non-i.i.d. data, using replay-based strategies to retain previous task knowledge. To address limitations in existing variations, which only select samples based on individual properties, a new metric will be introduced to evaluate and choose the optimal replay batch from the memory buffer. This will address the issue of selecting a set of seemingly optimal individual samples, but not the best set as a whole.
Keywords: continual learning, experience replay, data complexity, feature selection, batch attention, anomaly detection