Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Decomposable and Debiased Representations via Attribute-Centric Information Bottlenecks

Jinyung Hong, Eun Som Jeon, Changhoon Kim, Keun Hee Park, Utkarsh Nath, Yezhou Yang, Pavan Turaga, Theodore P. Pavlic

TL;DR

A novel debiasing framework is proposed, Debiasing Global Workspace, introducing attention-based information bottlenecks for learning compositional representations of attributes without defining specific bias types.

Abstract

Biased attributes, spuriously correlated with target labels in a dataset, can problematically lead to neural networks that learn improper shortcuts for classifications and limit their capabilities for out-of-distribution (OOD) generalization. Although many debiasing approaches have been proposed to ensure correct predictions from biased datasets, few studies have considered learning latent embedding consisting of intrinsic and biased attributes that contribute to improved performance and explain how the model pays attention to attributes. In this paper, we propose a novel debiasing framework, Debiasing Global Workspace, introducing attention-based information bottlenecks for learning compositional representations of attributes without defining specific bias types. Based on our observation that learning shape-centric representation helps robust performance on OOD datasets, we adopt those abilities to learn robust and generalizable representations of decomposable latent embeddings corresponding to intrinsic and biasing attributes. We conduct comprehensive evaluations on biased datasets, along with both quantitative and qualitative analyses, to showcase our approach's efficacy in attribute-centric representation learning and its ability to differentiate between intrinsic and bias-related features.

Learning Decomposable and Debiased Representations via Attribute-Centric Information Bottlenecks

TL;DR

A novel debiasing framework is proposed, Debiasing Global Workspace, introducing attention-based information bottlenecks for learning compositional representations of attributes without defining specific bias types.

Abstract

Biased attributes, spuriously correlated with target labels in a dataset, can problematically lead to neural networks that learn improper shortcuts for classifications and limit their capabilities for out-of-distribution (OOD) generalization. Although many debiasing approaches have been proposed to ensure correct predictions from biased datasets, few studies have considered learning latent embedding consisting of intrinsic and biased attributes that contribute to improved performance and explain how the model pays attention to attributes. In this paper, we propose a novel debiasing framework, Debiasing Global Workspace, introducing attention-based information bottlenecks for learning compositional representations of attributes without defining specific bias types. Based on our observation that learning shape-centric representation helps robust performance on OOD datasets, we adopt those abilities to learn robust and generalizable representations of decomposable latent embeddings corresponding to intrinsic and biasing attributes. We conduct comprehensive evaluations on biased datasets, along with both quantitative and qualitative analyses, to showcase our approach's efficacy in attribute-centric representation learning and its ability to differentiate between intrinsic and bias-related features.
Paper Structure (60 sections, 3 equations, 15 figures, 3 tables, 2 algorithms)

This paper contains 60 sections, 3 equations, 15 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Importance of Interpretable Concept Learning
  3. Intrinsically Interpretable Models in XAI
  4. Concept-Centric Transformers.
  5. Analysis on Out-of-Distribution Datasets
  6. Our Hypothesis.
  7. OOD Datasets.
  8. Experimental Results on OOD Datasets.
  9. Proposed Method
  10. Learning Latent Embedding of Attributes
  11. Produce Updated Attribute Representations
  12. Mixup between Original and Updated Representations
  13. Training Strategies
  14. Re-weighting Samples.
  15. Bias Attributes Augmentation and Scheduling.
  16. ...and 45 more sections

Figures (15)

  • Figure 1: CCT Activation Visualizations. (a) illustrates CCT's identification of bird body contours. (b) highlights the hen's crest contours. Both demonstrate CCT's concept focus without ground truth concept data.
  • Figure 2: Out-of-distribution (OOD) Benchmarking.
  • Figure 3: Visualization of classification preferences between shape and texture in images with conflicting cues. Human observers are indicated by red diamonds, various ImageNet-trained CNNs, including multiple ResNet architectures, by grey dots, and the CCT module by green dots. The CCT's results, represented by green dots, demonstrate a higher tendency to prioritize shape over texture, outperforming both standard and more complex ResNet variants, highlighting CCT's efficacy in shape-based recognition despite its relative parameter efficiency. The format of the figure is generated by geirhos2021partial.
  • Figure 4: Overall architecture of Debiasing Global Workspace (DGW)
  • Figure 5: Visualization of attention masks generated by intrinsic and bias components of DGW for the C-MNIST dataset
  • ...and 10 more figures