Case studies in the use of a hyperplane animator for neural network research

Neural network researchers can quantitatively examine several aspects of networks during training, such as changes in training set error, generalization error, and weights. However, a visual tool is often more appropriate for developing hypotheses about network learning behavior. When developing new neural network algorithms, insights can often be gained by visualizing the behavior of two-input networks geometrically; later the new method may be evaluated on higher dimensional problems. This paper presents case studies in which the animation of hyperplanes illustrated several new principles that govern neural network learning dynamics, and so led to new algorithms for network skeletonization, transfer, and training with positive examples only