Interpretable Deep Learning

IDL

The big data classification problems evolved from the scalability constraints may be addressed by the Deep Learning (DL) techniques. However, its performance may be adversely affected by the incorrect starting values of the connector and classifier parameters that are progressively modified through training and validation. If the starting values are selected randomly, close to zero, then a DL model starts with a new linear model and then progressively transforms into a nonlinear model. If the starting values are exactly zero, then the model does not evolve to a solution. Additionally, the mathematical connection between the connector and the classifier has never been incorporated in the DL models. In this paper, a perceptually inspired deep learning framework is proposed in which the edge sharpening filters and their frequency responses are used for the classifier and the connector parameters of a DL model to preserve class characteristics and regularize the DL parameters.