TL;DR
This paper introduces SWAG, a novel method for generating more accurate and interpretable explanations of Convolutional Neural Networks (CNNs) by weighting superpixels with guided gradients. SWAG offers a more efficient alternative to existing methods like LIME and RISE, achieving competitive local and global accuracy with significantly fewer computational passes.
What Problem Does This Paper Address?
As CNNs are increasingly deployed in critical applications like medical imaging and autonomous driving, understanding why they make certain decisions is paramount. Saliency maps are a popular tool for this, but there's a fundamental trade-off: explanations that precisely pinpoint important pixels (high accuracy) are often hard for humans to interpret, while visually intuitive "cognitive chunks" (like large regions) can be less accurate. Existing methods either use overly coarse regions (e.g., CAM-based methods) or individual pixels, or they require an excessive number of forward passes through the network to generate explanations (e.g., LIME, RISE). This paper aims to bridge this gap by creating explanations that are both accurate and interpretable, using a more efficient approach.
Methods
The core contribution is the SWAG (Superpixels Weighted by Average Gradients) method. It combines two key ideas:
- Superpixels as Cognitive Chunks: Instead of individual pixels or large feature maps, SWAG uses superpixels (groups of pixels that share visual characteristics) as the basic units for explanations. This provides a more interpretable level of granularity than single pixels.
- Gradient-Based Weighting: The importance of each superpixel is determined by the average of the backpropagated guided gradients within that superpixel. Guided backpropagation is used as it was found to perform slightly better than standard backpropagation.
The paper also proposes novel ways to generate these superpixels themselves, moving beyond traditional methods that only consider color and spatial proximity. They introduce two gradient-based superpixel generation methods:
- SWAG I+G: Superpixels are generated by clustering pixels based on both image color/spatial information and the guided backpropagated gradients.
- SWAG G: Superpixels are generated using only the guided backpropagated gradients.
This gradient-informed superpixel generation aims to create regions that better align with the CNN's internal understanding of the image, rather than just visual boundaries.
Datasets
The methods were evaluated on:
- ImageNet: A large-scale, general-purpose image dataset.
- Fine-grained datasets: Caltech-UCSD Birds 200 (CUB200), Stanford Dogs, and Oxford Flowers 102. These datasets are challenging because discriminative features are often small and localized.
Key Insights and Ideas
- Superpixels + Gradients = Better Explanations: The combination of using superpixels as interpretable regions and weighting them with guided gradients is a powerful approach that improves both accuracy and interpretability.
- Gradient-Aware Superpixel Generation is Crucial: Simply using standard superpixels isn't enough. Incorporating gradient information into the superpixel generation process (SWAG I+G and SWAG G) leads to explanations that better capture the network's decision-making process.
- Efficiency Matters: Methods like LIME and RISE, while capable, are computationally prohibitive due to their reliance on thousands of network passes. SWAG achieves competitive results with only a single forward and backward pass, making it much more practical.
- Local vs. Global Accuracy Trade-off: The paper highlights that different metrics (like deletion for local accuracy and ROAR for global accuracy) reveal different aspects of explanation quality. SWAG performs well on both, unlike some other methods that excel at one but not the other.
Key Results
- Local Accuracy: SWAG and its variants (SWAG I+G, SWAGG) consistently achieve better local accuracy (measured by deletion AUC) than Grad-CAM, Grad-CAM++, LIME, RISE, and XRAI across multiple datasets and models, often by a significant margin. SWAGG (gradient-only superpixels) generally performed best.
- Global Accuracy (ROAR): SWAG I+G showed the best performance among the proposed methods for global accuracy, outperforming other region-based techniques on CUB200 and Stanford Dogs.
- Efficiency: SWAG is significantly more efficient than LIME, RISE, and XRAI, requiring only a fraction of the computation time (seconds vs. minutes/hours). It is only slightly slower than Grad-CAM/Grad-CAM++.
- Weak-Localization: For the VGG16 network, SWAG achieved better weak-localization scores than Grad-CAM based methods.
Limitations and Future Work
- Stanford Dogs with ResNet50 Anomaly: For the Stanford Dogs dataset with ResNet50, LIME performed better than SWAG. The authors hypothesize this is because LIME's superpixel count better matched the discriminative regions (e.g., a dog's head) compared to SWAG's tendency to focus on smaller details like eyes and mouth.
- Hyperparameter Tuning for ROAR: The authors note that tuning hyperparameters for the ROAR (global accuracy) metric is impractical, and their current tuning was optimized for local interpretability. This suggests potential for further improvement in global accuracy.
- Alternative Pixel Scoring: For sanity checks, using methods other than guided backpropagation (like random noise or Sobel edges) to weight superpixels resulted in poorer local accuracy, reinforcing the importance of guided gradients.
- Future Work: The paper suggests that using alternative methods for pixel scoring (beyond guided backpropagation) could potentially yield even better results, especially for weak-localization. They also propose that their gradient-based superpixel generation methods could be used as a drop-in replacement for other superpixel-based explanation techniques like LIME.
Summary updated