Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Common-Sense Bias Modeling for Classification Tasks

Miao Zhang, Zee fryer, Ben Colman, Ali Shahriyari, Gaurav Bharaj

TL;DR

This paper introduces CSBD, a description-based framework to uncover and mitigate dataset biases in image classification. By clustering noun-phrase embeddings from image captions into common-sense features and measuring their dataset-wide co-occurrence with the target via the $cphie$ statistic, CSBD reveals a broad spectrum of potential biases beyond readily identifiable cues. A human-in-the-loop can filter spurious correlations, and bias mitigation is achieved through data re-weighting that decorrelates the presence of spurious features from the target, without requiring sensitive group labels. Experiments on CelebA-Dialog, MS-COCO 2014, and LVLM-generated captions demonstrate that CSBD discovers novel biases and achieves state-of-the-art debiasing performance while preserving classification quality. The approach emphasizes robustness to caption quality and shows generalizability across description sources, offering a practical pathway for bias auditing and mitigation in vision datasets.

Abstract

Machine learning model bias can arise from dataset composition: correlated sensitive features can distort the downstream classification model's decision boundary and lead to performance differences along these features. Existing de-biasing works tackle the most prominent bias features, such as colors of digits or background of animals. However, real-world datasets often include a large number of feature correlations that intrinsically manifest in the data as common sense information. Such spurious visual cues can further reduce model robustness. Thus, domain practitioners desire a comprehensive understanding of correlations and the flexibility to address relevant biases. To this end, we propose a novel framework to extract comprehensive biases in image datasets based on textual descriptions, a common sense-rich modality. Specifically, features are constructed by clustering noun phrase embeddings with similar semantics. The presence of each feature across the dataset is inferred, and their co-occurrence statistics are measured, with spurious correlations optionally examined by a human-in-the-loop module. Downstream experiments show that our method uncovers novel model biases in multiple image benchmark datasets. Furthermore, the discovered bias can be mitigated by simple data re-weighting to de-correlate the features, outperforming state-of-the-art unsupervised bias mitigation methods.

Common-Sense Bias Modeling for Classification Tasks

TL;DR

This paper introduces CSBD, a description-based framework to uncover and mitigate dataset biases in image classification. By clustering noun-phrase embeddings from image captions into common-sense features and measuring their dataset-wide co-occurrence with the target via the statistic, CSBD reveals a broad spectrum of potential biases beyond readily identifiable cues. A human-in-the-loop can filter spurious correlations, and bias mitigation is achieved through data re-weighting that decorrelates the presence of spurious features from the target, without requiring sensitive group labels. Experiments on CelebA-Dialog, MS-COCO 2014, and LVLM-generated captions demonstrate that CSBD discovers novel biases and achieves state-of-the-art debiasing performance while preserving classification quality. The approach emphasizes robustness to caption quality and shows generalizability across description sources, offering a practical pathway for bias auditing and mitigation in vision datasets.

Abstract

Machine learning model bias can arise from dataset composition: correlated sensitive features can distort the downstream classification model's decision boundary and lead to performance differences along these features. Existing de-biasing works tackle the most prominent bias features, such as colors of digits or background of animals. However, real-world datasets often include a large number of feature correlations that intrinsically manifest in the data as common sense information. Such spurious visual cues can further reduce model robustness. Thus, domain practitioners desire a comprehensive understanding of correlations and the flexibility to address relevant biases. To this end, we propose a novel framework to extract comprehensive biases in image datasets based on textual descriptions, a common sense-rich modality. Specifically, features are constructed by clustering noun phrase embeddings with similar semantics. The presence of each feature across the dataset is inferred, and their co-occurrence statistics are measured, with spurious correlations optionally examined by a human-in-the-loop module. Downstream experiments show that our method uncovers novel model biases in multiple image benchmark datasets. Furthermore, the discovered bias can be mitigated by simple data re-weighting to de-correlate the features, outperforming state-of-the-art unsupervised bias mitigation methods.
Paper Structure (24 sections, 3 equations, 4 figures, 7 tables, 1 algorithm)

This paper contains 24 sections, 3 equations, 4 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Unsupervised bias discovery.
  4. Bias mitigation.
  5. Bias discovery with human knowledge.
  6. Method
  7. Experiments
  8. Datasets
  9. CelebA-Dialog
  10. MS-COCO 2014
  11. LVLM-generated caption.
  12. Implementation
  13. Results
  14. Correlation discovery results
  15. Correlation robustness evaluation
  16. ...and 9 more sections

Figures (4)

  • Figure 1: Spurious features are everywhere: Objects which co-occur frequently with the target can affect model prediction, for example, kites and people in MS-COCO. Spurious features like this are of multiple types and may cause different downstream biases. Our method aims to discover a comprehensive list of them based on common sense descriptions, and treat biases which have not been explored in literature.
  • Figure 2: Common Sense Bias Discovery (CSBD) system and the mitigation strategy overview: (Left) Input corpus of image and corresponding description pairs are given. (Middle) (1) Descriptions are split into semantically meaningful noun chunks and encoded into representation vectors. (2) A hierarchical clustering on the vector set generates a list of common sense feature clusters. (3) Correlations between the presence of every two features are computed. (4) Highly-correlated features are examined by a human for mitigation. (Right) Finally, correlations are mitigated by adjusting image sampling weights during model training, automatically calculated from the derived feature presence statistics.
  • Figure 3: Part of correlated features and their correlation coefficients analyzed by CSBD on CelebA and MS-COCO datasets. Two types of common sense descriptions are used: human-generated (ground truth) and LVLM-generated caption.
  • Figure 4: T-SNE results for sample embeddings of cat classifiers ("Couch" as sensitive label) trained with MS-COCO. Target label: "Cat" (blue & orange) and "No cat" (green & red) are most linearly separable (model decision boundary is indicated by the dotted straight line) with CSBD compared to other baselines. For bias discovery, methods are marked with their assigned bias-aligning and bias-conflict samples (Pseudo group). Only CSBD discovers thus treats the sensitive feature couch (E).