Hyperparameter Tuning of ConvLSTM Network Models

Vrskova, R.; Sykora, P.; Kamencay, P.; Hudec, R.; Radil, R.

Deep learning algorithms have achieved amazing performance in computer vision area. However, a biggest problem deep learning has, is the high dependency on hyper-parameters. The algorithm results may be different, depending on hyper-parameters. This paper presents an effective method for hyper-parameter tuning using deep learning. The deep neural network structure for video classification using Convolutional Long Short-Term Memory (ConvLSTM) was used. The proposed method for hyper-parameter tuning using ConvLSTM was described. This proposed method with hyper-parameter tuning methods (Grid search, Bayesian optimization and Genetic algorithm) was compared. The experiment results show that proposed approach using ConvLSTM can be compared with the results obtained from the methods analogs to the proposed approach. However, we are looking for other hyper-parameters, for example number of filters, filter size, number of epochs, batch size and training optimization algorithm. The proposed approach can be used for correct or incorrect use of face mask during COVID-19 pandemic. © 2021 IEEE.