Extraction of fibre network architecture by X-ray tomography and prediction of elastic properties using an affine analytical model

This paper proposes a method for extracting reliable architectural characteristics from complex porous structures using micro-computed tomography (μCT) images. The work focuses on a highly porous material composed of a network of fibres bonded together. The segmentation process, allowing separation of the fibres from the remainder of the image, is the most critical step in constructing an accurate representation of the network architecture. Segmentation methods, based on local and global thresholding, were investigated and evaluated by a quantitative comparison of the architectural parameters they yielded, such as the fibre orientation and segment length (sections between joints) distributions and the number of inter-fibre crossings. To improve segmentation accuracy, a deconvolution algorithm was proposed to restore the original images. The efficacy of the proposed method was verified by comparing μCT network architectural characteristics with those obtained using high resolution CT scans (nanoCT). The results indicate that this approach resolves the architecture of these complex networks and produces results approaching the quality of nanoCT scans. The extracted architectural parameters were used in conjunction with an affine analytical model to predict the axial and transverse stiffnesses of the fibre network. Transverse stiffness predictions were compared with experimentally measured values obtained by vibration testing. © 2011 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.

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.

Paraphrastic language models and combination with neural network language models

In natural languages multiple word sequences can represent the same underlying meaning. Only modelling the observed surface word sequence can result in poor context coverage, for example, when using n-gram language models (LM). To handle this issue, paraphrastic LMs were proposed in previous research and successfully applied to a US English conversational telephone speech transcription task. In order to exploit the complementary characteristics of paraphrastic LMs and neural network LMs (NNLM), the combination between the two is investigated in this paper. To investigate paraphrastic LMs' generalization ability to other languages, experiments are conducted on a Mandarin Chinese broadcast speech transcription task. Using a paraphrastic multi-level LM modelling both word and phrase sequences, significant error rate reductions of 0.9% absolute (9% relative) and 0.5% absolute (5% relative) were obtained over the baseline n-gram and NNLM systems respectively, after a combination with word and phrase level NNLMs. © 2013 IEEE.

Modelling nonlinear vehicle dynamics with neural networks

The nonlinear modelling ability of neural networks has been widely recognised as an effective tool to identify and control dynamic systems, with applications including nonlinear vehicle dynamics which this paper focuses on using multi-layer perceptron networks. Existing neural network literature does not detail some of the factors which effect neural network nonlinear modelling ability. This paper investigates into and concludes on required network size, structure and initial weights, considering results for networks of converged weights. The paper also presents an online training method and an error measure representing the network's parallel modelling ability over a range of operating conditions. Copyright © 2010 Inderscience Enterprises Ltd.

Application of neural networks to the ranking of perinatal variables influencing birthweight.

In this paper we compare Multi-Layer Perceptrons (a neural network type) with Multivariate Linear Regression in predicting birthweight from nine perinatal variables which are thought to be related. Results show, that seven of the nine variables, i.e., gestational age, mother's body-mass index (BMI), sex of the baby, mother's height, smoking, parity and gravidity, are related to birthweight. We found no significant relationship between birthweight and each of the two variables, i.e., maternal age and social class.

Development of low entropy coding in a recurrent network.

In this paper we present an unsupervised neural network which exhibits competition between units via inhibitory feedback. The operation is such as to minimize reconstruction error, both for individual patterns, and over the entire training set. A key difference from networks which perform principal components analysis, or one of its variants, is the ability to converge to non-orthogonal weight values. We discuss the network's operation in relation to the twin goals of maximizing information transfer and minimizing code entropy, and show how the assignment of prior probabilities to network outputs can help to reduce entropy. We present results from two binary coding problems, and from experiments with image coding.

Dynamic learning with the EM algorithm for neural networks

In this paper, we derive an EM algorithm for nonlinear state space models. We use it to estimate jointly the neural network weights, the model uncertainty and the noise in the data. In the E-step we apply a forwardbackward Rauch-Tung-Striebel smoother to compute the network weights. For the M-step, we derive expressions to compute the model uncertainty and the measurement noise. We find that the method is intrinsically very powerful, simple and stable.

Introduction to the development of an information management system for soil mix technology using artificial neural networks

This paper introduces current work in collating data from different projects using soil mix technology and establishing trends using artificial neural networks (ANNs). Variation in unconfined compressive strength as a function of selected soil mix variables (e.g., initial soil water content and binder dosage) is observed through the data compiled from completed and on-going soil mixing projects around the world. The potential and feasibility of ANNs in developing predictive models, which take into account a large number of variables, is discussed. The main objective of the work is the management and effective utilization of salient variables and the development of predictive models useful for soil mix technology design. Based on the observed success in the predictions made, this paper suggests that neural network analysis for the prediction of properties of soil mix systems is feasible. © ASCE 2011.

An example of the use of neural computing techniques in materials science - The modelling of fatigue thresholds in Ni-base superalloys

Two adaptive numerical modelling techniques have been applied to prediction of fatigue thresholds in Ni-base superalloys. A Bayesian neural network and a neurofuzzy network have been compared, both of which have the ability to automatically adjust the network's complexity to the current dataset. In both cases, despite inevitable data restrictions, threshold values have been modelled with some degree of success. However, it is argued in this paper that the neurofuzzy modelling approach offers real benefits over the use of a classical neural network as the mathematical complexity of the relationships can be restricted to allow for the paucity of data, and the linguistic fuzzy rules produced allow assessment of the model without extensive interrogation and examination using a hypothetical dataset. The additive neurofuzzy network structure means that redundant inputs can be excluded from the model and simple sub-networks produced which represent global output trends. Both of these aspects are important for final verification and validation of the information extracted from the numerical data. In some situations neurofuzzy networks may require less data to produce a stable solution, and may be easier to verify in the light of existing physical understanding because of the production of transparent linguistic rules. © 1999 Elsevier Science S.A.

Preliminary model development for predicting strength and stiffness of cement-stabilized soils using artificial neural networks

This paper presents ongoing work on data collection and collation from a large number of laboratory cement-stabilization projects worldwide. The aim is to employ Artificial Neural Networks (ANN) to establish relationships between variables, which define the properties of cement-stabilized soils, and the two parameters determined by the Unconfined Compression Test, the Unconfined Compressive Strength (UCS), and stiffness, using E50 calculated from UCS results. Bayesian predictive neural network models are developed to predict the UCS values of cement-stabilized inorganic clays/silts, as well as sands as a function of selected soil mix variables, such as grain size distribution, water content, cement content and curing time. A model which can predict the stiffness values of cement-stabilized clays/silts is also developed and compared to the UCS model. The UCS model results emulate known trends better and provide more accurate estimates than the results from the E50 stiffness model. © 2013 American Society of Civil Engineers.

Avoiding pathologies in very deep networks

Choosing appropriate architectures and regularization strategies of deep networks is crucial to good predictive performance. To shed light on this problem, we analyze the analogous problem of constructing useful priors on compositions of functions. Specifically, we study the deep Gaussian process, a type of infinitely-wide, deep neural network. We show that in standard architectures, the representational capacity of the network tends to capture fewer degrees of freedom as the number of layers increases, retaining only a single degree of freedom in the limit. We propose an alternate network architecture which does not suffer from this pathology. We also examine deep covariance functions, obtained by composing infinitely many feature transforms. Lastly, we characterize the class of models obtained by performing dropout on Gaussian processes.