Learning word dependencies in text by means of a deep recurrent belief network

Publication Type:
Journal Article
Knowledge-Based Systems, 2016, 108 pp. 144 - 154
Issue Date:
Full metadata record
Files in This Item:
Filename Description Size
1-s2.0-S0950705116302349-main.pdfPublished Version719.3 kB
Adobe PDF
We propose a deep recurrent belief network with distributed time delays for learning multivariate Gaussians. Learning long time delays in deep belief networks is difficult due to the problem of vanishing or exploding gradients with increase in delay. To mitigate this problem and improve the transparency of learning time-delays, we introduce the use of Gaussian networks with time-delays to initialize the weights of each hidden neuron. From our knowledge of time delays, it is possible to learn the long delays from short delays in a hierarchical manner. In contrast to previous works, here dynamic Gaussian Bayesian networks over training samples are evolved using Markov Chain Monte Carlo to determine the initial weights of each hidden layer of neurons. In this way, the time-delayed network motifs of increasing Markov order across layers can be modeled hierarchically using a deep model. To validate the proposed Variable-order Belief Network (VBN) framework, it is applied for modeling word dependencies in text. To explore the generality of VBN, it is further considered for a real-world scenario where the dynamic movements of basketball players are modeled. Experimental results obtained showed that the proposed VBN could achieve over 30% improvement in accuracy on real-world scenarios compared to the state-of-the-art baseline
Please use this identifier to cite or link to this item: