Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Overfocusing Bias of Convolutional Neural Networks: A Saliency-Guided Regularization Approach

David Bertoin, Eduardo Hugo Sanchez, Mehdi Zouitine, Emmanuel Rachelson

TL;DR

CNNs often overfocus on narrow image regions in low-data regimes, harming generalization. The authors propose Saliency-Guided Dropout (SGDrop), which uses attribution maps to drop the most salient latent features via a $q_\rho$-quantile mask, encouraging broader attention; an EMA variant stabilizes mask computation. Empirically, SGDrop broadens attribution maps, increases neuron coverage, aligns saliency with labels and human explanations, and improves generalization across diverse datasets and architectures, with notable gains on VGG16 and ConvNeXt and beneficial transfer-learning effects. The approach is modular, attribution-method agnostic, and extensible, offering a practical regularization pathway that links interpretability insights with robustness in low-data settings.

Abstract

Despite transformers being considered as the new standard in computer vision, convolutional neural networks (CNNs) still outperform them in low-data regimes. Nonetheless, CNNs often make decisions based on narrow, specific regions of input images, especially when training data is limited. This behavior can severely compromise the model's generalization capabilities, making it disproportionately dependent on certain features that might not represent the broader context of images. While the conditions leading to this phenomenon remain elusive, the primary intent of this article is to shed light on this observed behavior of neural networks. Our research endeavors to prioritize comprehensive insight and to outline an initial response to this phenomenon. In line with this, we introduce Saliency Guided Dropout (SGDrop), a pioneering regularization approach tailored to address this specific issue. SGDrop utilizes attribution methods on the feature map to identify and then reduce the influence of the most salient features during training. This process encourages the network to diversify its attention and not focus solely on specific standout areas. Our experiments across several visual classification benchmarks validate SGDrop's role in enhancing generalization. Significantly, models incorporating SGDrop display more expansive attributions and neural activity, offering a more comprehensive view of input images in contrast to their traditionally trained counterparts.

The Overfocusing Bias of Convolutional Neural Networks: A Saliency-Guided Regularization Approach

TL;DR

CNNs often overfocus on narrow image regions in low-data regimes, harming generalization. The authors propose Saliency-Guided Dropout (SGDrop), which uses attribution maps to drop the most salient latent features via a -quantile mask, encouraging broader attention; an EMA variant stabilizes mask computation. Empirically, SGDrop broadens attribution maps, increases neuron coverage, aligns saliency with labels and human explanations, and improves generalization across diverse datasets and architectures, with notable gains on VGG16 and ConvNeXt and beneficial transfer-learning effects. The approach is modular, attribution-method agnostic, and extensible, offering a practical regularization pathway that links interpretability insights with robustness in low-data settings.

Abstract

Despite transformers being considered as the new standard in computer vision, convolutional neural networks (CNNs) still outperform them in low-data regimes. Nonetheless, CNNs often make decisions based on narrow, specific regions of input images, especially when training data is limited. This behavior can severely compromise the model's generalization capabilities, making it disproportionately dependent on certain features that might not represent the broader context of images. While the conditions leading to this phenomenon remain elusive, the primary intent of this article is to shed light on this observed behavior of neural networks. Our research endeavors to prioritize comprehensive insight and to outline an initial response to this phenomenon. In line with this, we introduce Saliency Guided Dropout (SGDrop), a pioneering regularization approach tailored to address this specific issue. SGDrop utilizes attribution methods on the feature map to identify and then reduce the influence of the most salient features during training. This process encourages the network to diversify its attention and not focus solely on specific standout areas. Our experiments across several visual classification benchmarks validate SGDrop's role in enhancing generalization. Significantly, models incorporating SGDrop display more expansive attributions and neural activity, offering a more comprehensive view of input images in contrast to their traditionally trained counterparts.
Paper Structure (17 sections, 2 equations, 4 figures, 11 tables)

This paper contains 17 sections, 2 equations, 4 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Saliency-Guided Dropout
  4. Generic algorithm
  5. Practical implementation
  6. Experiments
  7. Back to the preliminary experiment
  8. Saliencies and neural coverage
  9. Generalization performance
  10. Transfer learning
  11. Scheduling $\rho$
  12. Is the EMA useful?
  13. Computation overhead
  14. Limitations and perspectives
  15. Conclusion
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Overfocusing phenomenon in neural networks: rapid concentration on small areas of images.
  • Figure 2: Overview of our proposed method SGDrop.
  • Figure 3: Variation in attribution area ratios and associated saliency maps during training.
  • Figure 4: Qualitative comparison of saliency map for vanilla, dropout, and SGDrop training. Ground truth bounding boxes in green and attribution-derived bounding boxes in red.