Extended Few-Shot Learning: Exploiting Existing Resources for Novel Tasks

In many practical few-shot learning problems, even though labeled examples\nare scarce, there are abundant auxiliary datasets that potentially contain\nuseful information. We propose the problem of extended few-shot learning to\nstudy these scenarios. We then introduce a framework to address the challenges\nof efficiently selecting and effectively using auxiliary data in few-shot image\nclassification. Given a large auxiliary dataset and a notion of semantic\nsimilarity among classes, we automatically select pseudo shots, which are\nlabeled examples from other classes related to the target task. We show that\nnaive approaches, such as (1) modeling these additional examples the same as\nthe target task examples or (2) using them to learn features via transfer\nlearning, only increase accuracy by a modest amount. Instead, we propose a\nmasking module that adjusts the features of auxiliary data to be more similar\nto those of the target classes. We show that this masking module performs\nbetter than naively modeling the support examples and transfer learning by 4.68\nand 6.03 percentage points, respectively.\n