ORACLE: Order Robust Adaptive Continual Learning
Jaehong Yoon, Saehoon Kim, Eunho Yang and Sung Ju Hwang
The order of the tasks a continual learning model encounters may have large impact on the performance of each task, as well as the task-average performance. This order-sensitivity may cause serious problems in real-world scenarios where fairness plays a critical role (e.g. medical diagnosis). To tackle this problem, we propose a novel order-robust continual learning method, which instead of learning a completely shared set of weights, represent the parameters for each task as a sum of task-shared parameters that captures generic representations and task-adaptive parameters capturing task-specific ones, where the latter is factorized into sparse low-rank matrices in order to minimize capacity increase. With such parameter decomposition, when training for a new task, the task-adaptive parameters for earlier tasks remain mostly unaffected, where we update them only to reflect the changes made to the task-shared parameters. This prevents catastrophic forgetting for old tasks and at the same time make the model less sensitive to the task arrival order. We validate our Order-Robust Adaptive Continual LEarning (ORACLE) method on multiple benchmark datasets against state-of-the-art continual learning methods, and the results show that it largely outperforms those strong baselines with significantly less increase in capacity and training time, as well as obtains smaller performance disparity for each task with different order sequences.