Adi Hayat, Mark Kliger, Shachar Fleishman, Daniel Cohen-Or

Feb 15, 2018 (modified: Oct 27, 2017) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: Conventional deep learning classifiers are static in the sense that they are trained on a predefined set of classes and learning to classify a novel class typically requires re-training. In this work, we address the problem of Low-shot network-expansion learning. We introduce a learning framework which enables expanding a pre-trained (base) deep network to classify novel classes when the number of examples for the novel classes is particularly small. We present a simple yet powerful distillation method where the base network is augmented with additional weights to classify the novel classes, while keeping the weights of the base network unchanged. We term this learning hard distillation, since we preserve the response of the network on the old classes to be equal in both the base and the expanded network. We show that since only a small number of weights needs to be trained, the hard distillation excels for low-shot training scenarios. Furthermore, hard distillation avoids detriment to classification performance on the base classes. Finally, we show that low-shot network expansion can be done with a very small memory footprint by using a compact generative model of the base classes training data with only a negligible degradation relative to learning with the full training set.
  • TL;DR: In this paper, we address the problem of Low-shot network-expansion learning
  • Keywords: Low-Shot Learning, class incremental learning, Network expansion, Generative model, Distillation
0 Replies