Delve: Neural Network Feature Variance Analysis.

Designing neural networks is a complex task. Deep neural networks are often referred to as “black box” models little insight in the function they approximate is gained from looking at the structure of layer outputs. Delve is a tool for looking at how a neural network represents data, and how these representations, or features, change throughout training. This tool enables deep learning researchers to understand the limitations and suggest improvements for the design of their networks, such as removing or adding layers.