Bi-Objective Continual Learning: Learning ‘New’ While Consolidating ‘Known’

• DOI: 10.1609/aaai.v34i04.6060
• Publication Date: 2020-06-29T20:54:07Z
• Authors: Xiaoyu Tao; Xiaopeng Hong; Xinyuan Chang; Yihong Gong
• Keywords: Pillar; Consolidation

In this paper, we propose a novel single-task continual learning framework named Bi-Objective Continual Learning (BOCL). BOCL aims at both consolidating historical knowledge and from new data. On one hand, to preserve the old using small set of pillars, develop pillar consolidation (PLC) loss alleviate catastrophic forgetting problem. other contrastive (CPL) term improve classification performance, examine several data sampling strategies for efficient onsite ‘new’ with reasonable amount computational resources. Comprehensive experiments on CIFAR10/100, CORe50 subset ImageNet validate framework. We also reveal performance accuracy different when used finetune given CNN model. The code will be released.