Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ConvNets and ImageNet Beyond Accuracy: Understanding Mistakes and Uncovering Biases

Pierre Stock, Moustapha Cisse

TL;DR

The paper argues that image classification benchmarks and model assessments based on raw accuracy underestimate the true end-user experience, due to misalignments with human judgment and hidden biases. By combining human subject studies, feature- and example-based explanations, and adversarial-perturbation-based model criticism, it demonstrates that (a) ImageNet accuracy/robustness are underappreciated, (b) explanations help users interpret and trust predictions even for adversarial inputs, and (c) biases learned by models can be exposed and quantified. The findings show substantial reductions in rectified error when misclassifications are judged by humans, improved end-user agreement with adversarial predictions when explanations are provided, and effective uncovering of undesirable biases through adversarial model criticism. These results advocate for broader use of explanations and bias-detection tools to build more reliable, trustworthy vision systems and to motivate the development of richer benchmarks beyond single-label accuracy.

Abstract

ConvNets and Imagenet have driven the recent success of deep learning for image classification. However, the marked slowdown in performance improvement combined with the lack of robustness of neural networks to adversarial examples and their tendency to exhibit undesirable biases question the reliability of these methods. This work investigates these questions from the perspective of the end-user by using human subject studies and explanations. The contribution of this study is threefold. We first experimentally demonstrate that the accuracy and robustness of ConvNets measured on Imagenet are vastly underestimated. Next, we show that explanations can mitigate the impact of misclassified adversarial examples from the perspective of the end-user. We finally introduce a novel tool for uncovering the undesirable biases learned by a model. These contributions also show that explanations are a valuable tool both for improving our understanding of ConvNets' predictions and for designing more reliable models.

ConvNets and ImageNet Beyond Accuracy: Understanding Mistakes and Uncovering Biases

TL;DR

The paper argues that image classification benchmarks and model assessments based on raw accuracy underestimate the true end-user experience, due to misalignments with human judgment and hidden biases. By combining human subject studies, feature- and example-based explanations, and adversarial-perturbation-based model criticism, it demonstrates that (a) ImageNet accuracy/robustness are underappreciated, (b) explanations help users interpret and trust predictions even for adversarial inputs, and (c) biases learned by models can be exposed and quantified. The findings show substantial reductions in rectified error when misclassifications are judged by humans, improved end-user agreement with adversarial predictions when explanations are provided, and effective uncovering of undesirable biases through adversarial model criticism. These results advocate for broader use of explanations and bias-detection tools to build more reliable, trustworthy vision systems and to motivate the development of richer benchmarks beyond single-label accuracy.

Abstract

ConvNets and Imagenet have driven the recent success of deep learning for image classification. However, the marked slowdown in performance improvement combined with the lack of robustness of neural networks to adversarial examples and their tendency to exhibit undesirable biases question the reliability of these methods. This work investigates these questions from the perspective of the end-user by using human subject studies and explanations. The contribution of this study is threefold. We first experimentally demonstrate that the accuracy and robustness of ConvNets measured on Imagenet are vastly underestimated. Next, we show that explanations can mitigate the impact of misclassified adversarial examples from the perspective of the end-user. We finally introduce a novel tool for uncovering the undesirable biases learned by a model. These contributions also show that explanations are a valuable tool both for improving our understanding of ConvNets' predictions and for designing more reliable models.

Paper Structure

This paper contains 12 sections, 7 equations, 10 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Adversarial examples.
  4. Model Criticism.
  5. Feature-based Explanation.
  6. Experiments
  7. Human Subject Study of Classification Errors
  8. Study of Robustness to Adversarial Examples
  9. Adversarial Examples for Model Criticism
  10. Uncovering Biases with Model Criticism
  11. Remark
  12. Conclusion

Figures (10)

  • Figure 1: Top: Performance evolution of various CNN architectures on ImageNet. Bottom: Some images sampled from the Internet and misclassified by a ResNet-101.
  • Figure 2: Positive answers for misclassified samples. An image is prompted to 5 different subjects and a positive answer means the subject agrees with the predicted class
  • Figure 3: Some test samples misclassified by a ResNet-101 (first row) and a Densenet-161 (second row). The predicted class is indicated in red, the ground truth in black and in parenthesis. All those examples gathered more than four (4 or 5) positive answers over 5 on AMT. Note that no adversarial noise has been added to the images.
  • Figure 4: Positive answers for adversarial samples. The images are either displayed as a whole (entire image) or with attention (explanation). Every image is prompted to 5 different subjects. A positive answer means the subject agrees with the predicted adversarial class.
  • Figure 5: Left: an adversarial image of the true class Jeep predicted as Ambulance by the network. Center: the explanation of the clean image for its prediction (Jeep). Right: the explanation of the adversarial image for its prediction (Ambulance).
  • ...and 5 more figures