Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Robust Convolutional Neural Networks with Relevant Feature Focusing via Explanations

Kazuki Adachi, Shin'ya Yamaguchi

TL;DR

This work tackles distribution shift causing CNNs to latch onto spurious features by introducing ReFF, a training-time regularizer that aligns model explanations with task-relevant regions using Grad-CAM. A pseudo-annotator generates regional annotations to reduce manual labeling, enabling scalable application to off-the-shelf CNNs without extra test-time cost. Across Textured MNIST, ISIC2017, and Oxford Pets, ReFF improves test accuracy under varying spurious-feature regimes and benefits from multi-layer regularization, with pseudo-annotations often outperforming full manual annotations. The approach enhances robustness and explainability, offering a practical path to more reliable vision models in real-world, shift-prone environments.

Abstract

Existing image recognition techniques based on convolutional neural networks (CNNs) basically assume that the training and test datasets are sampled from i.i.d distributions. However, this assumption is easily broken in the real world because of the distribution shift that occurs when the co-occurrence relations between objects and backgrounds in input images change. Under this type of distribution shift, CNNs learn to focus on features that are not task-relevant, such as backgrounds from the training data, and degrade their accuracy on the test data. To tackle this problem, we propose relevant feature focusing (ReFF). ReFF detects task-relevant features and regularizes CNNs via explanation outputs (e.g., Grad-CAM). Since ReFF is composed of post-hoc explanation modules, it can be easily applied to off-the-shelf CNNs. Furthermore, ReFF requires no additional inference cost at test time because it is only used for regularization while training. We demonstrate that CNNs trained with ReFF focus on features relevant to the target task and that ReFF improves the test-time accuracy.

Learning Robust Convolutional Neural Networks with Relevant Feature Focusing via Explanations

TL;DR

This work tackles distribution shift causing CNNs to latch onto spurious features by introducing ReFF, a training-time regularizer that aligns model explanations with task-relevant regions using Grad-CAM. A pseudo-annotator generates regional annotations to reduce manual labeling, enabling scalable application to off-the-shelf CNNs without extra test-time cost. Across Textured MNIST, ISIC2017, and Oxford Pets, ReFF improves test accuracy under varying spurious-feature regimes and benefits from multi-layer regularization, with pseudo-annotations often outperforming full manual annotations. The approach enhances robustness and explainability, offering a practical path to more reliable vision models in real-world, shift-prone environments.

Abstract

Existing image recognition techniques based on convolutional neural networks (CNNs) basically assume that the training and test datasets are sampled from i.i.d distributions. However, this assumption is easily broken in the real world because of the distribution shift that occurs when the co-occurrence relations between objects and backgrounds in input images change. Under this type of distribution shift, CNNs learn to focus on features that are not task-relevant, such as backgrounds from the training data, and degrade their accuracy on the test data. To tackle this problem, we propose relevant feature focusing (ReFF). ReFF detects task-relevant features and regularizes CNNs via explanation outputs (e.g., Grad-CAM). Since ReFF is composed of post-hoc explanation modules, it can be easily applied to off-the-shelf CNNs. Furthermore, ReFF requires no additional inference cost at test time because it is only used for regularization while training. We demonstrate that CNNs trained with ReFF focus on features relevant to the target task and that ReFF improves the test-time accuracy.
Paper Structure (12 sections, 3 equations, 3 figures, 3 tables)

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

Table of Contents

  1. Introduction
  2. Related Work
  3. Focusing on Task-relevant Feature
  4. Explanation Regularizer
  5. Pseudo-annotator
  6. Experiments
  7. Datasets
  8. Training Details
  9. Effect of Spurious Features
  10. Effect of the Size of the Task-Relevant Features
  11. Effect of Adding the Pseudo-annotator
  12. Conclusion

Figures (3)

  • Figure 1: Example of spurious features. CNNs mistakenly learn the features of grassy fields to be 'sheep' because the training images labeled 'sheep' tend to contain grassy fields. But at test time, images of non-sheep in grassy fields are classified as 'sheep' and vice versa. The images are sampled from the AwA dataset animals-with-attributes.
  • Figure 2: Overview of relevant feature focusing (ReFF).
  • Figure 3: Grad-CAMs computed from the outputs of layer4. The red regions indicate where the classifiers attach importance in order to make a prediction, the blue regions are the opposite. The upper, middle, and lower rows show the inputs, baseline, and ReFF, respectively.