Class Activation Maps

Class Activation Maps


In neuroscience, a receptive field is the portion of sensory space to which a neuron fires if an appropriate stimulus is presented. Each neuron, for example, in (layer 5 of) the primary visual cortex responds to the orientation of a bar. By mapping the responses across layer 5 to a stimulus and mapping the responses to one layeer 5 neuron across all stimuli, we can understand the relationship between input and neural activity.

The analogous concept in computational neural networks is a class activation map. Computational neural networks have a related term, discriminative regions, that they borrowed from machine learning, which corresponds to the just noticeable difference in neuroscience.

Prior Work

There are posts and articles on how to visualize class activation maps for images. In this post I discuss how to visualize the class activation maps for molecules represented by a SMILES string.


To Read:

  1. Interpretation and Visualization Techniques for Deep Learning Models in Medical Imaging
  2. Discovering Molecular Functional Groups Using Graph Convolutional Networks