Introduction to Class Activation Maps

Class activation maps (CAMs)

Class activation maps (CAMs) explain the predictions made by a convolutional neural network (CNN). They are generated by the CNN and are used to visualize the regions of the image that are important for predicting a specific class. They are class-discriminative saliency maps that help distinguish between classes.

The main difference between saliency maps and CAMs is how they generate the explanation. Saliency maps generate the explanation by computing the gradient of the predicted class with respect to the input image. In contrast, CAMs generate the explanation by a linear combination of the feature maps from the last convolutional layer of a neural network. The nature of linear combination weights differs from algorithm to algorithm.

To understand the idea of CAMs better, let’s assume that the prediction from our neural network $f(.)$ can be written as: