I was going through this paper, Tao, Xiaoyu, et al. “Bi-Objective Continual Learning: Learning 'New ’ While Consolidating ‘Known’.” AAAI. 2020. (https://aaai.org/ojs/index.php/AAAI/article/view/6060) where the focus is on NI, rather than the popular NC. In the experiments section, their exact Cifar10/100 dataset creation details are a bit unclear:
We further split the training images into 5 sessions, and apply different illumination and saturation for each subsequent session as new training patterns.
Did anyone find the implementation details of the above? If so, kindly share. Thanks!
Edit: The author has uploaded the dataset preparation code in their GitHub repo.