Source-linked AI summary
Grad-CAM: Why did you say that?
Ramprasaath R Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, Dhruv Batra
TL;DR
CNNs achieve strong vision performance but remain difficult to interpret, motivating visual explanations that reveal why predictions are made. The paper introduces Grad-CAM and Guided Grad-CAM, then applies and evaluates them across CNN-based models, including image captioning and VQA. The reported results show class-discriminative, high-resolution explanations that help humans assess model trustworthiness and align with occlusion-based importance.
Problem
Accurate CNN-based models are difficult to interpret, creating a need for transparent explanations of their predictions.
Method
Grad-CAM uses class-specific gradients to weight convolutional feature maps, while Guided Grad-CAM fuses its localization map with Guided Backpropagation for high-resolution explanations.
Results
Human studies found that localization-augmented visualizations discriminate between classes more accurately and better reveal classifier trustworthiness, with applications also demonstrated for image captioning and VQA.
Takeaways & Limitations
Grad-CAM provides a way to understand CNN-based models across architectures and tasks without requiring architectural changes or retraining.
Abstract
from arXiv · showhide
We propose a technique for making Convolutional Neural Network (CNN)-based models more transparent by visualizing input regions that are 'important' for predictions -- or visual explanations. Our approach, called Gradient-weighted Class Activation Mapping (Grad-CAM), uses class-specific gradient information to localize important regions. These localizations are combined with existing pixel-space visualizations to create a novel high-resolution and class-discriminative visualization called Guided Grad-CAM. These methods help better understand CNN-based models, including image captioning and visual question answering (VQA) models. We evaluate our visual explanations by measuring their ability to discriminate between classes, to inspire trust in humans, and their correlation with occlusion maps. Grad-CAM provides a new way to understand CNN-based models. We have released code, an online demo hosted on CloudCV, and a full version of this extended abstract.
1. Introduction
Grad-CAM addresses the interpretability gap in accurate CNNs by producing class-discriminative localization maps without architectural changes or retraining. Guided Grad-CAM combines these maps with high-resolution pixel-space visualizations to explain predictions across CNN-based tasks.
- Motivation: Accurate deep networks are difficult to interpret because their internal computations are not decomposable into intuitive components.The paper frames transparency as important for understanding failures and building trust in intelligent systems.
- Motivation: A good visual explanation should both localize the predicted category and capture fine-grained image detail.The paper identifies class discrimination and high resolution as complementary explanation goals.
- Grad-CAM: Grad-CAM uses class-specific gradients to weight convolutional feature maps, producing a coarse, normalized localization map followed by a ReLU.The gradient-derived weights capture feature-map importance for a target class or differentiable activation.
- Grad-CAM: Unlike CAM, Grad-CAM applies directly to CNN-based differentiable architectures, including networks with fully connected layers, without replacing layers or retraining.The method generalizes CAM beyond architectures requiring global-average-pooled features before the output.
- Guided Grad-CAM: Guided Grad-CAM pointwise-multiplies Grad-CAM with Guided Backpropagation to combine class discrimination with high-resolution detail.The resulting visualization localizes the relevant object while preserving fine details such as stripes.
3. Experiments
The experiments evaluate Guided Grad-CAM through human class discrimination, model-trust judgments, and agreement with occlusion-based explanations, then apply it to classification errors, image captioning, and VQA.
- Evaluating Visualizations: Human studies compare visualizations from AlexNet and VGG-16 while restricting analysis to instances where both models match the ground truth.This separates visualization efficacy from differences in model accuracy.
- Evaluating Visualizations: Guided Grad-CAM makes VGG-16 appear clearly more reliable than AlexNet, whereas Guided Backpropagation yields only a slight reliability advantage.Relative Reliability scores range from -2 to +2, with positive scores favoring VGG-16.
- Evaluating Visualizations: Guided Grad-CAM intensities correlate with image-occlusion effects, indicating greater faithfulness to the original model than Guided Backpropagation.Occlusion measures score changes after masking image patches, and the comparison uses rank correlation.
- Evaluating Visualizations: Guided Grad-CAM is evaluated for human class discrimination, trust judgments, and faithfulness to model behavior.Table 1 uses Human Classification Accuracy, Relative Reliability, and Rank Correlation w/ Occlusion.
- Analyzing Failure Modes for VGG-16: Guided Grad-CAM helps inspect VGG-16 classification failures by explaining ambiguous or seemingly unreasonable predictions.Examples include sandbar versus volcano and a predicted syringes class that becomes interpretable through visualization.
- Image Captioning: For image captioning, Grad-CAM localizes regions supporting generated words, including small kites and people, while excluding unmentioned objects.Caption-specific maps can also remain within specified bounding boxes without using bounding-box annotations.
- Visual Question Answering: For VQA, Grad-CAM explanations focus on image regions associated with answer choices such as different parts of a firehydrant.The examples cover baseline and ResNet-based hierarchical co-attention VQA models.
4. Conclusion
The paper presents Grad-CAM as a class-discriminative localization technique for CNN-based visual explanations and combines it with high-resolution visualizations. Human studies and examples indicate that the resulting explanations support class discrimination, trust assessment, and understanding across CNN-based models.
- Conclusion: Grad-CAM provides class-discriminative localization, while Guided Grad-CAM combines it with high-resolution visualizations for CNN-based models.The conclusion identifies this combination as the paper’s central proposed technique.
- Conclusion: Human studies find that localization-augmented visualizations better discriminate between classes and reveal classifier trustworthiness.The conclusion also points to applications in image captioning and VQA.
- Conclusion: The full paper extends the evaluation to ImageNet localization, VGG-16 failure modes, VQA-human-attention correlation, ablations, and additional examples.These analyses are listed as part of the full version beyond the examples in the abstract.