Neural network technologies are increasingly used in solving problems from various fields of industry, agriculture, medicine. The problems of optimizing the architecture and ANN hyperparameters focused on solving various classes of applied problems are caused by the need to improve the quality of deep ANN functioning. There are various methods for optimizing ANN hyperparameters, for example, using genetic algorithms, but they involve the creation of additional software. The «Keras Tuner» toolkit is known to optimize the procedures for the computer selection of the set of ANN hyperparameters. It is a toolkit for automated search for the optimal combination of network hyperparameters. In the «Keras Tuner» toolkit, you can choose mathematical methods such as random search, Bayesian optimization, Hyperband. In the numerical experiments carried out, 14 hyperparameters were varied: the number of blocks of convolutional layers and the filters forming them, the type of activation functions, the parameters of the “dropout” regulatory layers, and others. The investigated toolkit has demonstrated high efficiency of hyperparameter optimization while simultaneously varying several tens of parameters of the convolutional network. The calculation time for the studied ANN hyperparameters was several tens of hours, without the use of GPU graphics accelerators on the Colaboratory platform.
Computer Optimization of ANN Hyperparameters for Retrospective Information Processing
Lect. Notes in Networks, Syst.
International School on Neural Networks, Initiated by IIASS and EMFCSC ; 2022 ; St.Petersburg, Russia February 08, 2022 - February 10, 2022
Networked Control Systems for Connected and Automated Vehicles ; Kapitel : 72 ; 723-730
16.11.2022
8 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
British Library Conference Proceedings | 2016
|American Institute of Physics | 2023
|A Novel Method to Select Hyperparameters of the DBSCAN Algorithm for RADAR Applications
SAE Technical Papers | 2024
|