Learn how to use Explainable AI  to design and implement popular explanation algorithms.

Mastering Image Classification

Explainable AI is a set of tools and frameworks that helps us understand and interpret the internal logic behind the predictions made by a deep learning network. With this, we can generate insights into the behavior and working of a model to mitigate issues around it in the development phase. 


In this module, we will be introduced to popular Explainable AI algorithms such as smooth gradient, integrated gradient, LIME, class activation maps, counterfactual explanations, feature attributions, etc., for image classification networks such as MobileNet-V2 trained on large-scale datasets like ImageNet-1K.

By the end of this module, we will understand the need for Explainable AI and be able to design and implement popular explanation algorithms like saliency maps, class activation maps, counterfactual explanations, etc. We will be able to evaluate and quantify the quality of the neural network explanations via several interpretability metrics.

Interpreting Image Classifier Decisions

A class activation map relies on gradients, whereas a saliency map relies on the linear combination of penultimate layer features.

A class activation map relies on the linear combination of penultimate layer features, whereas a saliency map relies on gradients. 

A class activation map visualizes importance at the feature level, whereas a saliency map visualizes importance at the pixel level.

A class activation map visualizes importance at the pixel level, whereas a saliency map visualizes importance at the feature level.

**(Select all that apply.)** How is a class activation map different from a saliency map?

(Select all that apply.) How is a class activation map different from a saliency map?

**(Select all that apply.)** Which algorithms use gradients for generating a CAM?

(Select all that apply.) Which algorithms use gradients for generating a CAM?

**(Select all that apply.)** The X-GradCAM algorithm satisfies which axioms?

(Select all that apply.) The X-GradCAM algorithm satisfies which axioms?

    weight = model.classifier[1].weight.data[prediction.item()]

    weight = torch.autograd.grad(logits[prediction.item()], features)[0]
    weight = weight.mean([2,3])[0]

    U, S, VT = np.linalg.svd(features, full_matrices=True)

    weight = torch.autograd.grad(logits[prediction.item()], features)[0]
    f_weight = features/(torch.sum(features, dim=[2,3], keepdim=True) + 1e-6)
    weight = torch.sum(weight*f_weight, dim=[2,3])[0]

Which code block can be used for computing linear combination weights used in Eigen-CAM?

Which code block can be used for computing linear combination weights used in Eigen-CAM?

    cam = weighted_features.reshape(h, w)

    cam = cam - cam.min()
    cam = cam/cam.max()

    heatmap = cam*0.4 + 0.6*np.array(original_image)/255.

Which code block is used for post-processing CAM?

Which code block is used for post-processing CAM?

Quiz yourself on the concepts learned in this chapter.

Quiz on Class Activation Maps