Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Advancing Ante-Hoc Explainable Models through Generative Adversarial Networks

Tanmay Garg, Deepika Vemuri, Vineeth N Balasubramanian

TL;DR

The paper tackles the interpretability gap in visual classification by proposing an ante-hoc framework that attaches an unsupervised explanation generator to a classifier and couples it with a Generative Adversarial Network (GAN) to guide concept learning. The method jointly optimizes classification, reconstruction, concept fidelity, and adversarial realism through a loss L that combines $L_c$, $L_R$, $L_F$, and GAN terms, encouraging latent concepts to align with human-interpretable properties. It demonstrates improved accuracy and coherent concept activations on CIFAR-10/100 across multiple GAN variants and noise strategies, while enabling qualitative concept visualization. This approach advances trustworthy AI in perception tasks by integrating generative feedback into concept learning and highlights practical trade-offs in training efficiency versus explainability gains.

Abstract

This paper presents a novel concept learning framework for enhancing model interpretability and performance in visual classification tasks. Our approach appends an unsupervised explanation generator to the primary classifier network and makes use of adversarial training. During training, the explanation module is optimized to extract visual concepts from the classifier's latent representations, while the GAN-based module aims to discriminate images generated from concepts, from true images. This joint training scheme enables the model to implicitly align its internally learned concepts with human-interpretable visual properties. Comprehensive experiments demonstrate the robustness of our approach, while producing coherent concept activations. We analyse the learned concepts, showing their semantic concordance with object parts and visual attributes. We also study how perturbations in the adversarial training protocol impact both classification and concept acquisition. In summary, this work presents a significant step towards building inherently interpretable deep vision models with task-aligned concept representations - a key enabler for developing trustworthy AI for real-world perception tasks.

Advancing Ante-Hoc Explainable Models through Generative Adversarial Networks

TL;DR

The paper tackles the interpretability gap in visual classification by proposing an ante-hoc framework that attaches an unsupervised explanation generator to a classifier and couples it with a Generative Adversarial Network (GAN) to guide concept learning. The method jointly optimizes classification, reconstruction, concept fidelity, and adversarial realism through a loss L that combines , , , and GAN terms, encouraging latent concepts to align with human-interpretable properties. It demonstrates improved accuracy and coherent concept activations on CIFAR-10/100 across multiple GAN variants and noise strategies, while enabling qualitative concept visualization. This approach advances trustworthy AI in perception tasks by integrating generative feedback into concept learning and highlights practical trade-offs in training efficiency versus explainability gains.

Abstract

This paper presents a novel concept learning framework for enhancing model interpretability and performance in visual classification tasks. Our approach appends an unsupervised explanation generator to the primary classifier network and makes use of adversarial training. During training, the explanation module is optimized to extract visual concepts from the classifier's latent representations, while the GAN-based module aims to discriminate images generated from concepts, from true images. This joint training scheme enables the model to implicitly align its internally learned concepts with human-interpretable visual properties. Comprehensive experiments demonstrate the robustness of our approach, while producing coherent concept activations. We analyse the learned concepts, showing their semantic concordance with object parts and visual attributes. We also study how perturbations in the adversarial training protocol impact both classification and concept acquisition. In summary, this work presents a significant step towards building inherently interpretable deep vision models with task-aligned concept representations - a key enabler for developing trustworthy AI for real-world perception tasks.
Paper Structure (12 sections, 2 equations, 3 figures, 4 tables)

This paper contains 12 sections, 2 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Background
  4. Methodology
  5. Proposed Architecture
  6. Experiments and Results
  7. Datasets and Comparison Methods
  8. Comparative Analysis
  9. Implementation Details
  10. Observations
  11. Conclusion and Future Work
  12. Acknowledgements

Figures (3)

  • Figure 1: Overview of our Proposed Architecture. N is the number of classes, C is the number of concepts
  • Figure 2: Top 5 images for CIFAR10 that activate the learnt concepts using cGAN (VGG 11) DAN (B=32, S=10). Eg: Cpt 2 captures antlers, Cpt 1 captures the color white - here we see that activated images are from different classes (ship, car).
  • Figure 3: Top 5 images for CIFAR100 that activate the learnt concepts (10 concepts from a subset of 100) using cGAN (VGG 19) DAN (B=32, S=10). Eg: Cpt 5 corresponds to color pink, Cpt 13 corresponds to object in ocean.