EVET: Enhancing Visual Explanations of Deep Neural Networks Using Image Transformations

Youngrock Oh, Hyungsik Jung, Jeonghyung Park, Min Soo KimWACV 2021

TL;DR

This paper introduces EVET, a novel pipeline that enhances existing visual explanation methods for deep neural networks by applying image transformations. EVET refines explanations by identifying regions consistently highlighted across transformed inputs, leading to more faithful, localized, and stable visual explanations with minimal computational overhead.

What Problem Does This Paper Address?

Deep neural networks, despite their impressive performance, often operate as "black boxes," making it difficult to understand why they make certain predictions. Visual explanation methods aim to address this by highlighting important regions of an input image that influence the model's decision. However, existing methods can sometimes produce noisy, unstable, or overly coarse explanations. The challenge is to improve the reliability and accuracy of these visual explanations without significantly increasing computational cost or requiring fundamental changes to the underlying explanation techniques.

Methods

EVET is a general pipeline that can be applied to any existing visual explanation method. The core idea is that regions consistently deemed important across various transformations of an input image are more likely to be truly critical for the model's prediction. The EVET pipeline involves three main steps:

  1. Image Transformations and Explanation Generation: The original input image is transformed using a set of geometric transformations (e.g., horizontal flip, scale, rotate, shear). For each transformed image, the chosen explanation method is applied to generate an importance map. These importance maps are then inversely transformed back to the original image space.
  2. Importance Map Assessment: To filter out potentially unreliable explanations, EVET assesses the target class probability of the transformed image. If a transformation significantly reduces the model's confidence in the original prediction, its corresponding importance map is down-weighted or omitted.
  3. Final Explanation Synthesis: The filtered and inversely transformed importance maps are combined using a weighted sum. The weights are determined by the target class probabilities of the perturbed images. Additionally, EVET incorporates a "stability index" by considering the sample standard deviation of importance values for each pixel across the transformed inputs. Pixels with low standard deviations (indicating stable importance) are boosted, leading to a refined and more robust final explanation map.

Datasets

The paper evaluates EVET on the following datasets:

  • PASCAL VOC (2007 segmentation task test set)
  • COCO (2014 validation set)
  • ImageNet (2012 validation set)

For quantitative evaluation, 1000 randomly sampled images were used from COCO and ImageNet.

Key Insights and Ideas

  • Consistency is Key: The fundamental insight is that true importance is revealed through consistency. If a region is highlighted by an explanation method across various transformations of an image, it's a strong indicator of its genuine importance to the model's decision.
  • Leveraging Existing Methods: EVET is designed as a plug-and-play enhancement. It doesn't replace existing methods but rather augments them, making it highly practical for researchers and practitioners.
  • Stability as a Metric: The paper introduces a "stability index" (SI) to quantitatively measure how robust an explanation method is to image transformations. This provides a valuable new metric for evaluating explanation techniques.
  • Computational Efficiency Gain: Surprisingly, applying EVET to computationally cheap methods like Grad-CAM can yield results comparable to more expensive state-of-the-art methods, while significantly reducing computation time.

Key Results

  • Improved Faithfulness, Localization, and Stability: EVET consistently improves the performance of various explanation methods (Gradient, Grad-CAM, Grad-CAM++, Score-CAM) across all tested metrics. This is demonstrated by lower average drop, higher increase in confidence, and higher win rates in faithfulness evaluations. Localization is also improved, as measured by the energy-based pointing game.
  • Significant Speedup: EVET-applied Grad-CAM (Grad-CAM*) achieves performance comparable to Score-CAM but with a drastic reduction in execution time (over 90% faster on VOC, COCO, and ImageNet).
  • Enhanced Stability: The proposed stability index (SI) shows that EVET significantly increases the robustness of explanations to image transformations, making them more reliable.

Limitations and Future Work

  • Transformation Choice: While EVET uses common geometric transformations, the paper acknowledges that the choice of transformations can influence results. Future work could explore other types of image modifications beyond geometric transformations.
  • Combining Maps: The current method linearly combines inverted importance maps. Future research could investigate more sophisticated methods for combining these maps.
  • Computational Cost of Transformations: Although generally negligible, the computational cost of image transformations themselves could become a factor with a very large number of transformations or complex ones.

The authors suggest exploring alternative methods for modifying input images and combining inverted importance maps as promising future research directions.

Summary updated

BibTeX
@inproceedings{oh2021evet,
  author = {Youngrock Oh and Hyungsik Jung and Jeonghyung Park and Min Soo Kim},
  title = {{EVET: Enhancing Visual Explanations of Deep Neural Networks Using Image Transformations}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2021},
}
Figure