Probing neural networks with t-SNE, class-specific projections and a guided tour

TL;DR

This paper employs graphical visualization techniques like t-SNE and class-specific projections to analyze and interpret the internal representations of neural networks classifying images, revealing how data organization evolves through layers.

Contribution

It introduces the use of class-specific principal component analogues and guided tours for dynamic visualization of neural network data representations.

Findings

t-SNE plots show increasing data organization in deeper layers

Class-specific projections reveal class separation and data typicality

Guided tours enhance animated visualization of neural network features

Abstract

We use graphical methods to probe neural nets that classify images. Plots of t-SNE outputs at successive layers in a network reveal increasingly organized arrangement of the data points. They can also reveal how a network can diminish or even forget about within-class structure as the data proceeds through layers. We use class-specific analogues of principal components to visualize how succeeding layers separate the classes. These allow us to sort images from a given class from most typical to least typical (in the data) and they also serve as very useful projection coordinates for data visualization. We find them especially useful when defining versions guided tours for animated data visualization.