INCREMENTAL LEARNING IN NON-STATIONARY ENVIRONMENTS

Abstract

This work focuses on source-free unsupervised incremental learning methods. We present an in-depth study of different types of incremental learning problem settings - domain incremental learning, feature incremental learning, and task incremental learning. We propose incremental learning methods for the three problem settings to retain past task knowledge in non-stationary environments. We evaluate our work on incremental disease classification model (Alzheimer's disease) problems, computer vision problems, and sentiment analysis problems. The proposed methods were evaluated on their capability to contextualize an unsupervised target task without catastrophic forgetting.