Emotional Features with Sparse Autoencoders

This is a showcase of features extracted from a sparse autoencoder trained on 1 million reddit images. Features are ranked by delta frequency (within-class % - across-class %) and labeled with the most correlated emotion. The code can be found here. Hover over an image to see a heatmap of where the feature is activated (sorry, not all images are processed).