[2D5-GS-2] Machine learning: applications (1)

[2D5-GS-2-05] Impact of Replay Method for Class Difficulty on Incremental Scenarios in Class Incremental Learning of Continual Learning

〇Hirono Kawashima¹, Makoto Kawano², Jin Nakazawa¹ (1. Keio University, 2. University of Tokyo)

Keywords:Continual Learning , Deep Learning, Image Classification

The purpose of this paper is to advance the investigation of the relationship between classes in class incremental learning with respect to class difficulty and catastrophic interference. As a continual learning method, we use CC-Replay, which controls the capacity of core data considering class difficulty, to verify the performance of CC-Replay in three CIFAR100 scenarios in which the hierarchical structure of incremental classes is intentionally varied. In addition to CIFAR-Rand, in which the classes to be incremented are randomly divided into arbitrary steps as in normal evaluation experiments, the scenario CIFAR-Block, in which the additional classes are solidified by the same parent class, and the scenario CIFAR-Equal was defined. The results showed that cc-replay had the highest final accuracy in both scenarios, and the step average accuracy was similar. When the hierarchy of incremental classes was different, it was found that accuracy did not decrease as the number of steps increased, but rather accuracy improved for the next step more than for the first step.