Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

How Well Do Self-Supervised Methods Perform in Cross-Domain Few-Shot Learning?

Yiyi Zhang, Ying Zheng, Xiaogang Xu, Jun Wang

TL;DR

It comes as a surprise that even with shallow architectures or small training datasets, self-supervised methods can perform favorably compared to the existing SOTA methods, and it is found that representations extracted from self- supervised methods exhibit stronger robustness than the supervised method.

Abstract

Cross-domain few-shot learning (CDFSL) remains a largely unsolved problem in the area of computer vision, while self-supervised learning presents a promising solution. Both learning methods attempt to alleviate the dependency of deep networks on the requirement of large-scale labeled data. Although self-supervised methods have recently advanced dramatically, their utility on CDFSL is relatively unexplored. In this paper, we investigate the role of self-supervised representation learning in the context of CDFSL via a thorough evaluation of existing methods. It comes as a surprise that even with shallow architectures or small training datasets, self-supervised methods can perform favorably compared to the existing SOTA methods. Nevertheless, no single self-supervised approach dominates all datasets indicating that existing self-supervised methods are not universally applicable. In addition, we find that representations extracted from self-supervised methods exhibit stronger robustness than the supervised method. Intriguingly, whether self-supervised representations perform well on the source domain has little correlation with their applicability on the target domain. As part of our study, we conduct an objective measurement of the performance for six kinds of representative classifiers. The results suggest Prototypical Classifier as the standard evaluation recipe for CDFSL.

How Well Do Self-Supervised Methods Perform in Cross-Domain Few-Shot Learning?

TL;DR

It comes as a surprise that even with shallow architectures or small training datasets, self-supervised methods can perform favorably compared to the existing SOTA methods, and it is found that representations extracted from self- supervised methods exhibit stronger robustness than the supervised method.

Abstract

Cross-domain few-shot learning (CDFSL) remains a largely unsolved problem in the area of computer vision, while self-supervised learning presents a promising solution. Both learning methods attempt to alleviate the dependency of deep networks on the requirement of large-scale labeled data. Although self-supervised methods have recently advanced dramatically, their utility on CDFSL is relatively unexplored. In this paper, we investigate the role of self-supervised representation learning in the context of CDFSL via a thorough evaluation of existing methods. It comes as a surprise that even with shallow architectures or small training datasets, self-supervised methods can perform favorably compared to the existing SOTA methods. Nevertheless, no single self-supervised approach dominates all datasets indicating that existing self-supervised methods are not universally applicable. In addition, we find that representations extracted from self-supervised methods exhibit stronger robustness than the supervised method. Intriguingly, whether self-supervised representations perform well on the source domain has little correlation with their applicability on the target domain. As part of our study, we conduct an objective measurement of the performance for six kinds of representative classifiers. The results suggest Prototypical Classifier as the standard evaluation recipe for CDFSL.
Paper Structure (15 sections, 2 equations, 3 figures, 4 tables)

This paper contains 15 sections, 2 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Few-shot Learning (FSL)
  4. Self-supervised Learning (SSL)
  5. Cross-domain Few-shot Learning (CDFSL)
  6. Preliminaries
  7. Problem Formulation
  8. Evaluated Methods
  9. Experiments
  10. Benchmarks
  11. To what extent self-supervised features can transfer directly to uncharted domains?
  12. A good embedding is all you need?
  13. Which one is the most practically effective among the plethora of self-supervised methods?
  14. Discussion
  15. Conclusion

Figures (3)

  • Figure 1: t-SNE plots of trained embeddings on 20 novel classes from MiniImageNet and 10 classes from EuroSAT, one color represents one class. Training methods considered are the supervised learned ResNet50, InfoMin and SimSiam. All are trained from ImageNet and tested respectively on MiniImageNet and EuroSAT.
  • Figure 2: The Coefficient of Variation (C.V) is calculated by ${\sigma \over \mu }$, where $\sigma$ defines the standard deviation among classifiers, $\mu$ stands for the average results of classifiers. Image is plotted from results listed in Table \ref{['tab:tabname4']}.
  • Figure 3: (a) the distance between the performance (%) of each classifier with their averaged performance; (b) The Z-Score of the averaged performance of each classifier among all classifiers. The performance of each classifier is firstly averaged by all methods in each dataset listed in Table \ref{['tab:tabname4']}.