Visualizing Latent Chemical Spaces




Visualizing Latent Chemical Spaces

Background

Class activation maps are an overlay of two heat maps. One represents the input to a convolutional neural network (CNN). The other, the extent to which activity in that part of the image contributes to activation of one unit in the final layer of the CNN. The meanings of the word activity and phrase part of the image are clear when the input and class activation map share the same natural representation. But, what if the input is a molecular structure, that is converted into a 2D matrix?

Objective:

Develop a class activation map extension for molecular structures.

  1. Class Activations Maps
  1. Analysis of Latent Spaces

Bibliography