Learning to extract task-adaptive representation for incremental few-shot learning

Abstract

Learning novel categories while preserving previously learned knowledge is a long-standing challenge in machine learning field. The challenge gets greater when a novel task is given with only a few labeled samples for each novel category, which is known as \textit{incremental few-shot learning} problem. We propose \textit{XtarNet}, which learns to extract task-adaptive representation for facilitating incremental few-shot learning. The method utilizes a backbone network pretrained on a set of base categories while also employing additional meta modules. These additional modules are trained across episodes via meta-learning based approach. Given a new task, the novel feature extracted from the meta-trained modules is combined with the base feature obtained from the pretrained model. The process of combining these two different features provides desired task-adaptive representation, which is called TAR. The TAR contains effective information to aid classifying both novel and base categories. The base and novel classifiers quickly adapt to a given task by utilizing the TAR. Experiments on standard image datasets indicate that our method achieves state-of-the-art incremental few-shot learning performance and further ablation shows the effectiveness of the TAR.