Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Data-Free Generative Replay for Class-Incremental Learning on Imbalanced Data

Sohaib Younis, Bernhard Seeger

TL;DR

DFGR tackles continual learning under class-incremental and imbalanced data by training a generator without real data, guided solely by batch-norm statistics and final-layer feature maps from a pre-trained classifier. The method combines a suite of losses for the generator (cross-entropy, BN statistics, feature maps, sample diversification, and image smoothing) with a classifier trained using focal loss and a dynamic replay adjustment, yielding balanced and diverse synthetic samples for rehearsal. Empirical results on MNIST and FashionMNIST show DFGR surpasses data-free baselines and highlights the pivotal role of cross-entropy loss and replay adjustment in achieving high accuracy while maintaining memory efficiency. The approach offers a practical pathway for data-free incremental learning in privacy-constrained or storage-limited settings and could extend to larger datasets in future work.

Abstract

Continual learning is a challenging problem in machine learning, especially for image classification tasks with imbalanced datasets. It becomes even more challenging when it involves learning new classes incrementally. One method for incremental class learning, addressing dataset imbalance, is rehearsal using previously stored data. In rehearsal-based methods, access to previous data is required for either training the classifier or the generator, but it may not be feasible due to storage, legal, or data access constraints. Although there are many rehearsal-free alternatives for class incremental learning, such as parameter or loss regularization, knowledge distillation, and dynamic architectures, they do not consistently achieve good results, especially on imbalanced data. This paper proposes a new approach called Data-Free Generative Replay (DFGR) for class incremental learning, where the generator is trained without access to real data. In addition, DFGR also addresses dataset imbalance in continual learning of an image classifier. Instead of using training data, DFGR trains a generator using mean and variance statistics of batch-norm and feature maps derived from a pre-trained classification model. The results of our experiments demonstrate that DFGR performs significantly better than other data-free methods and reveal the performance impact of specific parameter settings. DFGR achieves up to 88.5% and 46.6% accuracy on MNIST and FashionMNIST datasets, respectively. Our code is available at https://github.com/2younis/DFGR

Data-Free Generative Replay for Class-Incremental Learning on Imbalanced Data

TL;DR

DFGR tackles continual learning under class-incremental and imbalanced data by training a generator without real data, guided solely by batch-norm statistics and final-layer feature maps from a pre-trained classifier. The method combines a suite of losses for the generator (cross-entropy, BN statistics, feature maps, sample diversification, and image smoothing) with a classifier trained using focal loss and a dynamic replay adjustment, yielding balanced and diverse synthetic samples for rehearsal. Empirical results on MNIST and FashionMNIST show DFGR surpasses data-free baselines and highlights the pivotal role of cross-entropy loss and replay adjustment in achieving high accuracy while maintaining memory efficiency. The approach offers a practical pathway for data-free incremental learning in privacy-constrained or storage-limited settings and could extend to larger datasets in future work.

Abstract

Continual learning is a challenging problem in machine learning, especially for image classification tasks with imbalanced datasets. It becomes even more challenging when it involves learning new classes incrementally. One method for incremental class learning, addressing dataset imbalance, is rehearsal using previously stored data. In rehearsal-based methods, access to previous data is required for either training the classifier or the generator, but it may not be feasible due to storage, legal, or data access constraints. Although there are many rehearsal-free alternatives for class incremental learning, such as parameter or loss regularization, knowledge distillation, and dynamic architectures, they do not consistently achieve good results, especially on imbalanced data. This paper proposes a new approach called Data-Free Generative Replay (DFGR) for class incremental learning, where the generator is trained without access to real data. In addition, DFGR also addresses dataset imbalance in continual learning of an image classifier. Instead of using training data, DFGR trains a generator using mean and variance statistics of batch-norm and feature maps derived from a pre-trained classification model. The results of our experiments demonstrate that DFGR performs significantly better than other data-free methods and reveal the performance impact of specific parameter settings. DFGR achieves up to 88.5% and 46.6% accuracy on MNIST and FashionMNIST datasets, respectively. Our code is available at https://github.com/2younis/DFGR
Paper Structure (20 sections, 14 equations, 3 figures, 3 tables)

This paper contains 20 sections, 14 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Regularization-based Methods
  4. Knowledge distillation
  5. Overview of DFGR
  6. Methodology
  7. Generator Training
  8. Cross-entropy loss
  9. Batch-normalization statistics loss
  10. Feature map loss
  11. Sample diversification loss
  12. Image smoothing loss
  13. Total generator loss
  14. Classifier Training/Re-training
  15. Focal loss
  16. ...and 5 more sections

Figures (3)

  • Figure 1: Long tail distribution of the whole Pl@ntNet dataset (with PlantCLEF 2014 subset in red) joly2014species.
  • Figure 2: a) Workflow of the classifier training with real images (for the first task). b) Workflow for re-training the classifier with real and generated images.
  • Figure 3: Workflow for training the generator.