Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Strong Baseline for Semi-Supervised Incremental Few-Shot Learning

Linglan Zhao, Dashan Guo, Yunlu Xu, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, Xiangzhong Fang

TL;DR

A novel paradigm containing two parts is proposed: a well-designed meta-training algorithm for mitigating ambiguity between base and novel classes caused by unreliable pseudo labels and a model adaptation mechanism to learn discriminative features for novel classes while preserving base knowledge using few labeled and all the unlabeled data.

Abstract

Few-shot learning (FSL) aims to learn models that generalize to novel classes with limited training samples. Recent works advance FSL towards a scenario where unlabeled examples are also available and propose semi-supervised FSL methods. Another line of methods also cares about the performance of base classes in addition to the novel ones and thus establishes the incremental FSL scenario. In this paper, we generalize the above two under a more realistic yet complex setting, named by Semi-Supervised Incremental Few-Shot Learning (S2 I-FSL). To tackle the task, we propose a novel paradigm containing two parts: (1) a well-designed meta-training algorithm for mitigating ambiguity between base and novel classes caused by unreliable pseudo labels and (2) a model adaptation mechanism to learn discriminative features for novel classes while preserving base knowledge using few labeled and all the unlabeled data. Extensive experiments on standard FSL, semi-supervised FSL, incremental FSL, and the firstly built S2 I-FSL benchmarks demonstrate the effectiveness of our proposed method.

A Strong Baseline for Semi-Supervised Incremental Few-Shot Learning

TL;DR

A novel paradigm containing two parts is proposed: a well-designed meta-training algorithm for mitigating ambiguity between base and novel classes caused by unreliable pseudo labels and a model adaptation mechanism to learn discriminative features for novel classes while preserving base knowledge using few labeled and all the unlabeled data.

Abstract

Few-shot learning (FSL) aims to learn models that generalize to novel classes with limited training samples. Recent works advance FSL towards a scenario where unlabeled examples are also available and propose semi-supervised FSL methods. Another line of methods also cares about the performance of base classes in addition to the novel ones and thus establishes the incremental FSL scenario. In this paper, we generalize the above two under a more realistic yet complex setting, named by Semi-Supervised Incremental Few-Shot Learning (S2 I-FSL). To tackle the task, we propose a novel paradigm containing two parts: (1) a well-designed meta-training algorithm for mitigating ambiguity between base and novel classes caused by unreliable pseudo labels and (2) a model adaptation mechanism to learn discriminative features for novel classes while preserving base knowledge using few labeled and all the unlabeled data. Extensive experiments on standard FSL, semi-supervised FSL, incremental FSL, and the firstly built S2 I-FSL benchmarks demonstrate the effectiveness of our proposed method.

Paper Structure

This paper contains 15 sections, 8 equations, 4 figures, 5 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related works
  3. Few-Shot Learning
  4. Few-Shot Incremental Learning
  5. Semi-Supervised Few-Shot Learning
  6. Task Formulation
  7. Methodology
  8. Preliminary
  9. Prototype Refinement with Fake Unlabeled Data
  10. Model Adaptation without Forgetting
  11. Experiments
  12. Experiment Setup
  13. Experimental Results
  14. Ablation Studies
  15. Conclusion

Figures (4)

  • Figure 1: Comparisons between our proposed S$^2$I-FSL and other FSL benchmarks. $N_b$ and $N$ are the numbers of base and novel classes, respectively. Best viewed in color.
  • Figure 2: An overview of our proposed method for S$^2$I-FSL. The left and right parts illustrate the model training and testing stages, respectively. Best viewed in color.
  • Figure 3: T-SNE t-SNE plots of queries and prototypes from a tiered-ImageNet test episode with and without our proposed components. Categories are represented by different colors.
  • Figure 4: Performance of different methods by varying the sample ratio in unlabeled set.