Architecture Optimization Model for the Deep Neural Network For Binary Classification Problems

Abstract

The daunting and challenging tasks of specifying the optimal network architecture and its parameters are still a major area of research in the field of Machine Learning (ML) till date. These tasks though determine the success of building and training an effective and accurate model, are yet to be considered on a deep network having three hidden layers with varying optimized parameters to the best of our knowledge. This is due to expert’s opinion that it is practically difficult to determine a good Multilayer Perceptron (MLP) topology with more than two or three hidden layers without considering the number of samples and complexity of the classification to be learnt. In this study, a novel approach that combines an evolutionary genetic algorithm and an optimization algorithm and a supervised deep neural network (Deep-NN) using alternative activation functions with the view of modeling the prediction for the admission of prospective university students. The genetic algorithm is used to select optimal network parameters for the Deep-NN. Thus, this study presents a novel methodology that is effective, automatic and less human-dependent in finding optimal solution to diverse binary classification benchmarks. The model is trained, validated and tested using various performance metrics to measure the generalization ability and its performance.