Quadrant-distance graphs: a method for visualizing neural network weight spaces

Abstract

One of the major drawbacks to neural networks is the inability of the user to understand what is happening inside the network. Quadrant-distance (QD) graphs allow the user to graphically display a network's weight vector at any point in training, for networks of any size. This allows the user to quickly and easily identify similarities or differences between solution sets. QD graphs may also be used for a variety of other analysis functions, such as comparing initial weights to final weights, and observing the path of the network as it finds a solution.