Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhanced Few-Shot Class-Incremental Learning via Ensemble Models

Mingli Zhu, Zihao Zhu, Sihong Chen, Chen Chen, Baoyuan Wu

TL;DR

A new ensemble model framework cooperated with data augmentation to boost generalization is designed, which works as a library storing abundant features to guarantee fast adaptation to downstream tasks and mitigate the overfitting problem in FSCIL.

Abstract

Few-shot class-incremental learning (FSCIL) aims to continually fit new classes with limited training data, while maintaining the performance of previously learned classes. The main challenges are overfitting the rare new training samples and forgetting old classes. While catastrophic forgetting has been extensively studied, the overfitting problem has attracted less attention in FSCIL. To tackle overfitting challenge, we design a new ensemble model framework cooperated with data augmentation to boost generalization. In this way, the enhanced model works as a library storing abundant features to guarantee fast adaptation to downstream tasks. Specifically, the multi-input multi-output ensemble structure is applied with a spatial-aware data augmentation strategy, aiming at diversifying the feature extractor and alleviating overfitting in incremental sessions. Moreover, self-supervised learning is also integrated to further improve the model generalization. Comprehensive experimental results show that the proposed method can indeed mitigate the overfitting problem in FSCIL, and outperform the state-of-the-art methods.

Enhanced Few-Shot Class-Incremental Learning via Ensemble Models

TL;DR

A new ensemble model framework cooperated with data augmentation to boost generalization is designed, which works as a library storing abundant features to guarantee fast adaptation to downstream tasks and mitigate the overfitting problem in FSCIL.

Abstract

Few-shot class-incremental learning (FSCIL) aims to continually fit new classes with limited training data, while maintaining the performance of previously learned classes. The main challenges are overfitting the rare new training samples and forgetting old classes. While catastrophic forgetting has been extensively studied, the overfitting problem has attracted less attention in FSCIL. To tackle overfitting challenge, we design a new ensemble model framework cooperated with data augmentation to boost generalization. In this way, the enhanced model works as a library storing abundant features to guarantee fast adaptation to downstream tasks. Specifically, the multi-input multi-output ensemble structure is applied with a spatial-aware data augmentation strategy, aiming at diversifying the feature extractor and alleviating overfitting in incremental sessions. Moreover, self-supervised learning is also integrated to further improve the model generalization. Comprehensive experimental results show that the proposed method can indeed mitigate the overfitting problem in FSCIL, and outperform the state-of-the-art methods.
Paper Structure (22 sections, 5 equations, 6 figures, 8 tables)

This paper contains 22 sections, 5 equations, 6 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Few-shot learning and incremental learning.
  4. Few-shot class-incremental learning.
  5. Data augmentation.
  6. Methodology
  7. Problem formulation
  8. Overall framework
  9. Our method
  10. Multi-in multi-output ensemble model
  11. Integration with spatial-aware data augmentation
  12. Integration with self-supervision
  13. Experiments
  14. Experimental setup
  15. Comparison with the state-of-the-art
  16. ...and 7 more sections

Figures (6)

  • Figure 1: The Illustration of our method. It consists of three components: (1) Ensemble model structure. The multi-input multi-output network is used as the backbone. (2) Spatial-aware data augmentation. The background of images is augmented to alleviate overfitting problem for few-shot learning; (3) Self-supervision loss. A contrastive learning loss is used to boost the performance of our backbone.
  • Figure 2: Comparison of the proposed method with state-of-the-art on three benchmarks. Our method outperforms all the other methods.
  • Figure 3: Overfitting problem in normal training and ensemble learning scenarios.
  • Figure 4: Accuracy of base and new classes on CUB200 dataset.
  • Figure 5: Accuracy of each classifier and the whole model on miniImageNet.
  • ...and 1 more figures