Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unsupervised Learning for Class Distribution Mismatch

Pan Du, Wangbo Zhao, Xinai Lu, Nian Liu, Zhikai Li, Chaoyu Gong, Suyun Zhao, Hong Chen, Cuiping Li, Kai Wang, Yang You

TL;DR

This work proposes Unsupervised Learning for Class Distribution Mismatch (UCDM), which constructs positive-negative pairs from unlabeled data for classifier training and introduces a confidence-based labeling mechanism that iteratively assigns pseudo-labels to valuable real-world data and incorporates them into the training process.

Abstract

Class distribution mismatch (CDM) refers to the discrepancy between class distributions in training data and target tasks. Previous methods address this by designing classifiers to categorize classes known during training, while grouping unknown or new classes into an "other" category. However, they focus on semi-supervised scenarios and heavily rely on labeled data, limiting their applicability and performance. To address this, we propose Unsupervised Learning for Class Distribution Mismatch (UCDM), which constructs positive-negative pairs from unlabeled data for classifier training. Our approach randomly samples images and uses a diffusion model to add or erase semantic classes, synthesizing diverse training pairs. Additionally, we introduce a confidence-based labeling mechanism that iteratively assigns pseudo-labels to valuable real-world data and incorporates them into the training process. Extensive experiments on three datasets demonstrate UCDM's superiority over previous semi-supervised methods. Specifically, with a 60% mismatch proportion on Tiny-ImageNet dataset, our approach, without relying on labeled data, surpasses OpenMatch (with 40 labels per class) by 35.1%, 63.7%, and 72.5% in classifying known, unknown, and new classes.

Unsupervised Learning for Class Distribution Mismatch

TL;DR

This work proposes Unsupervised Learning for Class Distribution Mismatch (UCDM), which constructs positive-negative pairs from unlabeled data for classifier training and introduces a confidence-based labeling mechanism that iteratively assigns pseudo-labels to valuable real-world data and incorporates them into the training process.

Abstract

Class distribution mismatch (CDM) refers to the discrepancy between class distributions in training data and target tasks. Previous methods address this by designing classifiers to categorize classes known during training, while grouping unknown or new classes into an "other" category. However, they focus on semi-supervised scenarios and heavily rely on labeled data, limiting their applicability and performance. To address this, we propose Unsupervised Learning for Class Distribution Mismatch (UCDM), which constructs positive-negative pairs from unlabeled data for classifier training. Our approach randomly samples images and uses a diffusion model to add or erase semantic classes, synthesizing diverse training pairs. Additionally, we introduce a confidence-based labeling mechanism that iteratively assigns pseudo-labels to valuable real-world data and incorporates them into the training process. Extensive experiments on three datasets demonstrate UCDM's superiority over previous semi-supervised methods. Specifically, with a 60% mismatch proportion on Tiny-ImageNet dataset, our approach, without relying on labeled data, surpasses OpenMatch (with 40 labels per class) by 35.1%, 63.7%, and 72.5% in classifying known, unknown, and new classes.
Paper Structure (66 sections, 2 theorems, 40 equations, 21 figures, 17 tables, 2 algorithms)

This paper contains 66 sections, 2 theorems, 40 equations, 21 figures, 17 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. SSL under Class Distribution Mismatch
  4. Diffusion-Based Generation Methods
  5. Unsupervised Learning for CDM
  6. Preliminary
  7. Diffusion probabilistic models (DPMs)
  8. DPM's relationship to score-based generative models
  9. Problem Definition of UCDM
  10. Overview of training and testing datasets.
  11. Classifier architecture.
  12. Positive Instance Generation
  13. Negative Instance Generation
  14. Erasing semantic class via conditional inversion.
  15. Preserving visual characteristics in the unconditional reverse process.
  16. ...and 51 more sections

Key Result

Theorem 3.1

(Conditional DDIM song2020denoising inversion: progressive movement of the noise vector away from semantic class): Let $\bm{x}_t$ denote the noise vector at time step $t$ in the conditional inversion, and let $\mathcal{C}_y$ be the prompt of class $y$. Define $\delta_t = {\epsilon}_{\theta}(\bm{x}_t where $s_i \!=\!\!\sqrt{\alpha_t (1-{\bar{\alpha}}_{i+1})} \psi(\alpha_{i+1}, \alpha_{i}, 0)$ contr

Figures (21)

  • Figure 1: (a) Examples of SSL for closed-set task ($\text{SCDM}_{\text{CT}}$), open-set task (SCDM), and our proposed unsupervised learning for class distribution mismatch (UCDM), where no labels are used during training. (b) Accuracy of methods on closed-set and open-set tasks. In the closed-set task, samples are classified into known classes, while in the open-set task, they may be classified as unified "other" class, including both unknown and new categories.
  • Figure 2: Pipelines for instance generation. (a) and (b) show that the semantic class in the prompt can be synthesized in the positive instance pipeline or erased in the negative instance pipeline for a given seed sample. If the seed sample lacks the specified semantic class, the generated image resembles the original image.
  • Figure 3: The framework for training an unsupervised classifier based on generated positive and negative instances.
  • Figure 4: Ablation studies: (a) shows the ablation study on learning objectives, demonstrating the effectiveness of each component. (b) compares our method with SSL across varying label counts, highlighting its cost-saving potential. (c) analyzes the sensitivity to the confidence threshold, suggesting a higher threshold for stable performance.
  • Figure 5: Schematic diagram of loss functions.
  • ...and 16 more figures

Theorems & Definitions (5)

  • Theorem 3.1
  • Theorem 3.2
  • proof
  • proof
  • proof