Probing neural networks with t-SNE, class-specific projections and a guided tour
This work provides incremental visualization tools for researchers analyzing neural network behavior in image classification.
The authors tackled the problem of visualizing and understanding how neural networks process image data across layers, using t-SNE plots to show increasingly organized data arrangements and class-specific projections to reveal class separation and typicality rankings.
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.