The Effect of Varying Training on Neural Network Weights and Visualizations

Neural networks are used throughout modern society to solve many problems commonly thought of as impossible for computers. As their use becomes more widespread, an issue of measurement rises. In order to create metrics with which to measure the ability of neural networks, new techniques must be constantly developed. The purpose of this research is to determine whether varying training produces measurable and consistent patterns in the visualizations and weights of Convolutional Neural Networks. In order to carry out this investigation, a convolutional neural network was designed and run with varying levels of training to see if consistent, accurate, and precise changes or patterns could be observed. To determine if such a change or pattern existed, the weights and filters were analyzed through visualizations, both qualitatively and quantitatively. Several patterns were discovered in the visualizations, but they were inconsistent across layers and the quantitative models were only consistent in specific circumstances. This indicated that while training introduced and strengthened patterns in the weights and visualizations, the patterns observed may not be consistent between all neural networks.