Learning the Structure of Deep Sparse Graphical Models

Deep belief networks are a powerful way to model complex probability distributions. However, learning the structure of a belief network, particularly one with hidden units, is difficult. The Indian buffet process has been used as a nonparametric Bayesian prior on the directed structure of a belief network with a single infinitely wide hidden layer. In this paper, we introduce the cascading Indian buffet process (CIBP), which provides a nonparametric prior on the structure of a layered, directed belief network that is unbounded in both depth and width, yet allows tractable inference. We use the CIBP prior with the nonlinear Gaussian belief network so each unit can additionally vary its behavior between discrete and continuous representations. We provide Markov chain Monte Carlo algorithms for inference in these belief networks and explore the structures learned on several image data sets.

Competitive training in hidden Markov models

The use of hidden Markov models is placed in a connectionist framework, and an alternative approach to improving their ability to discriminate between classes is described. Using a network style of training, a measure of discrimination based on the a posteriori probability of state occupation is proposed, and the theory for its optimization using error back-propagation and gradient ascent is presented. The method is shown to be numerically well behaved, and results are presented which demonstrate that when using a simple threshold test on the probability of state occupation, the proposed optimization scheme leads to improved recognition performance.

A supply network configuration perspective on international supply chain development

Purpose - The purpose of this paper is to explore the impact of configuration on supply network capability. It was believed that a configuration perspective might provide new insights on the capability and performance of supply networks, a gap in the literature, and provide a basis for the development of tools to aid their analysis and design. Design/methodology/ approach - The methodology involved the development of a configuration definition and mapping approach extending established strategic and firm level constructs to the network operational level. The resulting tools were tested and refined in a series of case studies across a range of sectors and value chain models. Supply network capability assessments, from the perspective of the focal firm, were then compared with their configuration profiles. Findings - The configuration mapping tools were found to give new insights into the structure of supply networks and allow comparisons to be made across sectors and business models through the use of consistent and quantitative methods and common presentation. They provide the foundations for linking configuration to capability and performance, and contribute to supply network design and development by highlighting the intrinsic capabilities associated with different configurations. Research limitations/implications - Although multiple case networks have been investigated, the configuration exemplars remain suggestive models. The research suggests that a re-evaluation of operational process excellence models is needed, where the link between process maturity and performance may require a configuration context. Practical implications - Advantages of particular configurations have been identified with implications for supply network development and industrial policy. Originality/value - The paper seeks to develop established strategic management configuration concepts to the analysis and design of supply networks by providing a robust operational definition of supply network configuration and novel tools for their mapping and assessment. © Emerald Group Publishing Limited.

An Infinite Latent Attribute Model for Network Data

Latent variable models for network data extract a summary of the relational structure underlying an observed network. The simplest possible models subdivide nodes of the network into clusters; the probability of a link between any two nodes then depends only on their cluster assignment. Currently available models can be classified by whether clusters are disjoint or are allowed to overlap. These models can explain a "flat" clustering structure. Hierarchical Bayesian models provide a natural approach to capture more complex dependencies. We propose a model in which objects are characterised by a latent feature vector. Each feature is itself partitioned into disjoint groups (subclusters), corresponding to a second layer of hierarchy. In experimental comparisons, the model achieves significantly improved predictive performance on social and biological link prediction tasks. The results indicate that models with a single layer hierarchy over-simplify real networks.

Improving LVCSR system combination using neural network language model cross adaptation

State-of-the-art large vocabulary continuous speech recognition (LVCSR) systems often combine outputs from multiple subsystems developed at different sites. Cross system adaptation can be used as an alternative to direct hypothesis level combination schemes such as ROVER. The standard approach involves only cross adapting acoustic models. To fully exploit the complimentary features among sub-systems, language model (LM) cross adaptation techniques can be used. Previous research on multi-level n-gram LM cross adaptation is extended to further include the cross adaptation of neural network LMs in this paper. Using this improved LM cross adaptation framework, significant error rate gains of 4.0%-7.1% relative were obtained over acoustic model only cross adaptation when combining a range of Chinese LVCSR sub-systems used in the 2010 and 2011 DARPA GALE evaluations. Copyright © 2011 ISCA.

Using mesoscopic models to design strong and tough biomimetic polymer networks

Using computational modeling, we investigate the mechanical properties of polymeric materials composed of coiled chains, or "globules", which encompass a folded secondary structure and are cross-linked by labile bonds to form a macroscopic network. In the presence of an applied force, the globules can unfold into linear chains and thereby dissipate energy as the network is deformed; the latter attribute can contribute to the toughness of the material. Our goal is to determine how to tailor the labile intra- and intermolecular bonds within the network to produce material exhibiting both toughness and strength. Herein, we use the lattice spring model (LSM) to simulate the globules and the cross-linked network. We also utilize our modified Hierarchical Bell model (MHBM) to simulate the rupture and reforming of N parallel bonds. By applying a tensile deformation, we demonstrate that the mechanical properties of the system are sensitive to the values of N in and N out, the respective values of N for the intra- and intermolecular bonds. We find that the strength of the material is mainly controlled by the value of N out, with the higher value of N out providing a stronger material. We also find that, if N in is smaller than N out, the globules can unfold under the tensile load before the sample fractures and, in this manner, can increase the ductility of the sample. Our results provide effective strategies for exploiting relatively weak, labile interactions (e.g., hydrogen bonding or the thiol/disulfide exchange reaction) in both the intra- and intermolecular bonds to tailor the macroscopic performance of the materials. © 2011 American Chemical Society.

Development of predictive models for cement stabilized soils

Factors that affect the engineering properties of cement stabilized soils such as strength are discussed in this paper using data on these factors. The selected factors studied in this paper are initial soil water content, grain size distribution, organic matter content, binder dosage, age and curing temperature, which has been collated from a number of international deep mixing projects. Some resulting correlations from this data are discussed and presented. The concept of Artificial Neural Networks and its applicability in developing predictive models for deep mixed soils is presented and discussed using a subset of the collated data. The results from the neural network model were found to emulate the known trends and reasonable estimates of strength as a function of the selected variables were obtained. © 2012 American Society of Civil Engineers.