Online Hyperparameter Optimization for Class-Incremental Learning

Class-incremental learning (CIL) aims to train a classification model while the number of classes increases phase-by-phase. An inherent challenge CIL is stability-plasticity tradeoff, i.e., models should keep stable retain old knowledge and plastic absorb new knowledge. However, none existing can achieve optimal tradeoff in different data-receiving settings—where typically training-from-half (TFH) setting needs more stability, but training-from-scratch (TFS) plasticity. To this end, we design an online method that adaptively optimize without knowing as priori. Specifically, first introduce key hyperparameters influence e.g., distillation (KD) loss weights, rates, classifier types. Then, formulate hyperparameter optimization process Markov Decision Process (MDP) problem propose specific algorithm solve it. We apply local estimated rewards classic bandit Exp3 address issues when applying MDP methods protocol. Our consistently improves top-performing both TFH TFS settings, boosting average accuracy by 2.2 percentage points on ImageNet-Full, compared state-of-the-art. Code provided at https://class-il.mpi-inf.mpg.de/online/