Discovering the learned rules of dress collocation inside neural network mechanism

Abstract

This study is to capture the implicit rules of dress collocation by means of neural network modelling and analyses of the trained hidden structure. First, a multi-layer network model is adapted for training, where the input data are features designed by experiments to represent the various dressing styles of our selected nine fashion brands. Then we introduce a technique to display the inner categorization of the trained network model by a tree structure. From this, we discover the hidden rules of neural network models, and reveal the potential of local modification and correction without re-training the whole model.