Hyperparameter optimization for autoencoders that perform content-based image retrieval

Abstract

Content-based image retrieval problem is important in a wide range of computer science-based applications. One of the best solutions for this problem is the application of deep autoencoders. The accuracy of a neural network may be significantly dependent on the chosen of hyperparameters. In this work, we applied three methods of hyperparameter optimization (Tree of Parzen Estimator[TPE], annealing, and random search methods) to the recently proposed deep autoencoder architecture that solves the problem of image classification and content-based image retrieval. The parameters of AdamOptimizer were chosen as hyperparameters. Rather than a simple linear relationship between parameters and neural network accuracy, we observe significant difficulties in obtaining convergence with optimal hyperparameter values in all analyzed methods.