Course Home Page

Grad-CAM

This chapter covers results in the paper Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization available at this Arxiv link 2019 version.


Lecture

The lecture slides are available here.

Introduction to Grad-CAM

--- primaryColor: steelblue shuffleQuestions: false shuffleAnswers: true --- ### Grad-CAM has not been applied to which of the following in this paper? - [ ] CNNs with fully-connected layers. - [ ] CNNs for structured outputs such as captioning. - [ ] Large language models. > Correct. - [ ] CNNs for multi-modal inputs such as visual question answering.

Prior Results

--- primaryColor: steelblue shuffleQuestions: false shuffleAnswers: true --- ### An earlier approach CAM (Class Activation Maps) was not applicable to: - [ ] CNNs without fully-connected layers. - [x] CNNs with fully-connected layers, such as ResNet. > Correct. - [ ] CNNs without fully-connected layers but involving multiple layers of convolutions.

Grad-CAM: Technical Details

--- primaryColor: steelblue shuffleQuestions: false shuffleAnswers: true --- ### Grad-CAM is - [ ] independent of activation maps in the networks. - [x] the positive component of a weighted linear combination of activation maps. > Correct. - [ ] the positive component of an unweighted linear combination of activation maps. ### Guided Backpropagation and Grad-CAM visualizations are fused using: - [ ] the Fast Fourier Transform. - [x] elementwise multiplication. > Correct. - [ ] a two-layer neural network and optimization.

Grad-CAM as Generalization of CAM

Counterfactual Explanations

--- primaryColor: steelblue shuffleQuestions: false shuffleAnswers: true --- ### Counterfactuals are computed in this paper by using - [ ] random search over the input space. - [x] computing negations of the gradients. > Correct. - [ ] computing the attribution and removing them from the input.

Evaluation and Conclusions

--- primaryColor: steelblue shuffleQuestions: false shuffleAnswers: true --- ### The evaluation process in the paper does not address the question: - [ ] Given two attributions, which one is more trustworthy? - [x] employed a graph theory based algorithm developed earlier. > Correct. - [ ] Which of the two object categories is depicted in the (attribution) image?

Code and Assignment

There is no assignment for this lecture. The next few assignments will demonstrate and leverage concepts presented in this lecture.