Neural Network for Modeling Esthetic Selection

Some real world problems require significant human interaction for labeling the data, which is very expensive. Worse, in some cases, the exercise of human judgement is inherently subjective and contextual, and so the entire labeling must be done in one session, which may be too long. Our domain is the automatic generation of Mondrian-like images with an interactive interface for the user to select images. We use back-propagation neural networks to learn an approximation of a viewer's aesthetic using 2 category labelled data (images liked/disliked). We construct a data set for training in a sequential fashion related to the interactive art appreciation task, and produce an output profile which well approximates a regression task, even trained on classification data. Analysis of the learned network produces some surprises, with the discovery of some input contributions which are unexpected to the user.