Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LCGC: Learning from Consistency Gradient Conflicting for Class-Imbalanced Semi-Supervised Debiasing

Weiwei Xing, Yue Cheng, Hongzhu Yi, Xiaohui Gao, Xiang Wei, Xiaoyu Guo, Yuming Zhang, Xinyu Pang

TL;DR

A consistently conflicting gradient-based debiasing scheme dubbed LCGC is proposed, by encouraging biased class predictions during training to significantly improve the prediction accuracy of existing CISSL models on public benchmarks.

Abstract

Classifiers often learn to be biased corresponding to the class-imbalanced dataset, especially under the semi-supervised learning (SSL) set. While previous work tries to appropriately re-balance the classifiers by subtracting a class-irrelevant image's logit, but lacks a firm theoretical basis. We theoretically analyze why exploiting a baseline image can refine pseudo-labels and prove that the black image is the best choice. We also indicated that as the training process deepens, the pseudo-labels before and after refinement become closer. Based on this observation, we propose a debiasing scheme dubbed LCGC, which Learning from Consistency Gradient Conflicting, by encouraging biased class predictions during training. We intentionally update the pseudo-labels whose gradient conflicts with the debiased logits, representing the optimization direction offered by the over-imbalanced classifier predictions. Then, we debiased the predictions by subtracting the baseline image logits during testing. Extensive experiments demonstrate that LCGC can significantly improve the prediction accuracy of existing CISSL models on public benchmarks.

LCGC: Learning from Consistency Gradient Conflicting for Class-Imbalanced Semi-Supervised Debiasing

TL;DR

A consistently conflicting gradient-based debiasing scheme dubbed LCGC is proposed, by encouraging biased class predictions during training to significantly improve the prediction accuracy of existing CISSL models on public benchmarks.

Abstract

Classifiers often learn to be biased corresponding to the class-imbalanced dataset, especially under the semi-supervised learning (SSL) set. While previous work tries to appropriately re-balance the classifiers by subtracting a class-irrelevant image's logit, but lacks a firm theoretical basis. We theoretically analyze why exploiting a baseline image can refine pseudo-labels and prove that the black image is the best choice. We also indicated that as the training process deepens, the pseudo-labels before and after refinement become closer. Based on this observation, we propose a debiasing scheme dubbed LCGC, which Learning from Consistency Gradient Conflicting, by encouraging biased class predictions during training. We intentionally update the pseudo-labels whose gradient conflicts with the debiased logits, representing the optimization direction offered by the over-imbalanced classifier predictions. Then, we debiased the predictions by subtracting the baseline image logits during testing. Extensive experiments demonstrate that LCGC can significantly improve the prediction accuracy of existing CISSL models on public benchmarks.

Paper Structure

This paper contains 21 sections, 3 theorems, 17 equations, 5 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Classifier Debias
  4. Class-Imbalanced Semi-Supervised Learning
  5. Problem Setup and Preliminaries
  6. Problem Setup
  7. Backbone SSL Algorithm
  8. Class-Distribution-Mismatch-Aware Debiasing
  9. Integrated Gradients
  10. Why and How a Baseline Image Can Debias the Classifier Logits
  11. LCGC: A Consistency Gradient Conflicting Guided Debiasing Method
  12. Training a Consistency Conflicting Guided Model
  13. Refinement of Biased Logits During Inference
  14. Experiments
  15. Experimental Setup
  16. ...and 6 more sections

Key Result

Lemma 1

For a biased classifier trained on the class-imbalanced datasets, the basic CISSL model's logit $g_{\theta}(x_n)$ and its refinement $g’_\theta(x_n)$ have diverse predictions:

Figures (5)

  • Figure 1: Visualization of sensitive maps produced by the FixMatch model and CDMAD at the image. Left-to-right: original input image, visualization of sensitive maps produced by the FixMatch, CDMAD. The lightness of the sensitive map indicates how much attention the models pay to a particular area of the input image. The sensitive maps obtained by CDMAD have less noise.
  • Figure 2: The plot of $\mathcal{L}_{kl}(\theta)$ under CIFAR10-LT dataset and different imbalance ratios $\gamma_l$, $\gamma_u$. Both columns show that refined by the baseline image, the KL divergence value continues to decrease, indicating that the logit output distribution of the model before and after refinement becomes closer.
  • Figure 3: Pseudo-label refinement process using LCGC.
  • Figure 4: (a) and (b) present the class probabilities that take the absolute value of the log predicted on a black image using the proposed algorithm. (c) and (d) present the confusion matrices of the class predictions on test samples.
  • Figure 5: Line chart of validation bACC and GM for the CIFAR-10-LT ($\gamma_l = \gamma_u = 100$) dataset across a range of hyperparameter $\lambda$.

Theorems & Definitions (3)

  • Lemma 1
  • Theorem 1: Integrated gradient flow for class debiasing
  • Corollary 1.1