An Analysis of the Effects and Interaction of Hyper Parameters in Convolutional Neural Networks

Abstract

Deep Neural Network (DNN) hyper parameters are currently physically tuned, so robotized AI methods are being developed to find the best combination of hyper parameters. Finding the right design is difficult since the results of AutoML calculations heavily rely on the underlying frameworks. As a result, applying a visual analytics method is considered as a viable solution. As a solution, HyperTendril, an internet data discovery system is developed to enable user-driven hyperparameter tuning procedures in a model-independent configuration. HyperTendril encompasses a progressive approach to successfully execute the AutoML framework through a repetitive tuning process to assist clients in the tuning and execution of the AutoML framework based on a client's knowledge on the obtained outcome. Users can use HyperTendril to diagnose the configurations of various hyperparameter search methods and gain insights into their complex behaviour. Furthermore, HyperTendril provides various feature analysis to help users narrow their search areas based on the relative relevance of various hyperparameters and the consequences of their interactions.