When Does Self-supervision Improve Few-shot Learning?

  • 2019-10-08 17:47:14
  • Jong-Chyi Su, Subhransu Maji, Bharath Hariharan
  • 1


We present a technique to improve the generalization of deep representationslearned on small labeled datasets by introducing self-supervised tasks asauxiliary loss functions. Although recent research has shown benefits ofself-supervised learning (SSL) on large unlabeled datasets, its utility onsmall datasets is unknown. We find that SSL reduces the relative error rate offew-shot meta-learners by 4%-27%, even when the datasets are small and onlyutilizing images within the datasets. The improvements are greater when thetraining set is smaller or the task is more challenging. Though the benefits ofSSL may increase with larger training sets, we observe that SSL can have anegative impact on performance when there is a domain shift betweendistribution of images used for meta-learning and SSL. Based on this analysiswe present a technique that automatically select images for SSL from a large,generic pool of unlabeled images for a given dataset using a domain classifierthat provides further improvements. We present results using severalmeta-learners and self-supervised tasks across datasets with varying degrees ofdomain shifts and label sizes to characterize the effectiveness of SSL forfew-shot learning.


Quick Read (beta)

This feature is not avaialbe for this paper.