Title: A video prediction method by using long short-term memory-based adaptive structural learning of deep belief network and its investigation of input sequence length for data structure
Authors: Shin Kamada; Takumi Ichimura
Addresses: Advanced Artificial Intelligence Project Research Center, Research Organization of Regional Oriented Studies, Prefectural University of Hiroshima, 1-1-71, Ujina-Higashi, Minami-ku, Hiroshima 734-8558, Japan ' Advanced Artificial Intelligence Project Research Center, Prefectural University of Hiroshima, 1-1-71, Ujina-Higashi, Minami-ku, Hiroshima 734-8558, Japan; Faculty of Management and Information System, Prefectural University of Hiroshima, 1-1-71, Ujina-Higashi, Minami-ku, Hiroshima 734-8558, Japan
Abstract: In our research, the adaptive structural learning method using RBM and DBN has been developed as a deep learning model, which can find a suitable size of network structure for given input space during its training. The adaptive RBM and DBN were extended to the time-series prediction using the idea of long short-term memory (LSTM). Our previous research tackled the problems for supervised learning, in this paper, we challenge to reveal the power of our proposed method in the video recognition using Moving MNIST, which is a benchmark dataset for video recognition. Adaptive LSTM-DBN trained the time-series movement from the video data and it showed higher future prediction performance than the existing LSTM models (more than 90% for test data). Moreover, the detailed prediction results were investigated by training adaptive LSTM-DBN with various length of input sequence in the experiment results.
Keywords: deep learning; deep belief network; DBN; adaptive structural learning method; video recognition.
International Journal of Computational Intelligence Studies, 2021 Vol.10 No.2/3, pp.198 - 215
Received: 16 Apr 2020
Accepted: 13 Jul 2020
Published online: 19 May 2021 *