This paper presents the Tiled Computing Array (TCA), a simple, uniform, 3D-mesh packaging at inter-board level, for massively parallel computers. In particular, the power modelling and practical feasibility of the system is examined. TCA eliminates the need for hierarchical rackmount-structures and introduces short and immediate data channels in multiple physical orientations, allowing a more direct physical mapping of 3D computational topology to real hardware. A dedicated simulation platform has been developed, and an engineered prototype demonstrator has been built. This paper explores the feasibility of the TCA concept for current hardware technologies and systems, evaluates power modeling and validation, and highlights some of the novel design challenges associated with such a system. Evaluations of physical scalability toward large-scale systems are reported, showing that TCA is a promising approach.

In this paper, we introduce a theoretical basis for a Hadoop-based neural network for parallel and distributed feature selection in Big Data sets. It is underpinned by an associative memory (binary) neural network which is highly amenable to parallel and distributed processing and fits with the Hadoop paradigm. There are many feature selectors described in the literature which all have various strengths and weaknesses. We present the implementation details of five feature selection algorithms constructed using our artificial neural network framework embedded in Hadoop YARN. Hadoop allows parallel and distributed processing. Each feature selector can be divided into subtasks and the subtasks can then be processed in parallel. Multiple feature selectors can also be processed simultaneously (in parallel) allowing multiple feature selectors to be compared. We identify commonalities among the five features selectors. All can be processed in the framework using a single representation and the overall processing can also be greatly reduced by only processing the common aspects of the feature selectors once and propagating these aspects across all five feature selectors as necessary. This allows the best feature selector and the actual features to select to be identified for large and high dimensional data sets through exploiting the efficiency and flexibility of embedding the binary associative-memory neural network in Hadoop.

In this paper, a methodology for facial feature identification and localization approach is proposed based on binary neural network algorithms. We present a head pose and facial expression invariant 3D shape descriptor called Mesh-like Multi Circle Curvature Descriptor (MMCCD), which provides more 3D curvature attributes than other similar approaches. To search and match the feature patterns with more attributes, we use Advanced Uncertain Reasoning Architecture (AURA) k-Nearest Neighbour algorithms to encode, train and match the feature patterns based on 3D shape curvature. Experiments performed on the FRGC dataset (4950 3D faces) with pose and expression variations show that our approach is able to achieve an accurate (over 99.69% nose tip identification) and robust identification and localization of facial features.

Twenty years ago a paper detailing a novel neural network, called ADAM and that could be directly implemented in hardware RAM, was published in the first conference of this series. Subsequent research based directly/indirectly on this type of RAM-based neural network founded a research group that has produced over 200 research documents. This paper overviews that research and goes on to mathematically define a CMML, a generalised version of a CMM (the component at the heart of ADAM). The CMML can be trained to replicate the exact computational properties of a CMM and so is a plug-and-play replacement to a CMM; whilst a different training algorithm gives it different properties when used in recall.

We describe a high performance grid based signal search tool for distributed diagnostic applications developed in conjunction with Rolls-Royce plc for civil aero engine condition monitoring applications. With the introduction of advanced monitoring technology into engineering systems, healthcare, etc., the associated diagnostic processes are increasingly required to handle and consider vast amounts of data. An exemplar of such a diagnosis process was developed during the DAME project, which built a proof of concept demonstrator to assist in the enhanced diagnosis and prognosis of aero-engine conditions. In particular it has shown the utility of an interactive viewing and high performance distributed search tool (the signal data explorer) in the aeroengine diagnostic process. The viewing and search techniques are equally applicable to other domains. The signal data explorer and search services have been demonstrated on the Worldwide Universities Network to search distributed databases of electrocardiograph data.

A Grid based implementation of a system for finding duplicates in large databases is described. The solution is scalable to many nodes and does not suffer the problems found in other implementations that can result of loss of data and/or deadlock. The system may be applied to conventional de-duplication problems such as found in address management as well as more advanced problems such as banned image detection. The system uses the AURA pattern match methods implemented within a service oriented architecture. The approach builds on the PMS and PMC technology developed in the DAME eScience project.

A correlation matrix memory (CMM) is a form of binary neural network, that can be used for high-speed approximate search and match operations on large unstructured datasets. Typically, the processing requirements for a CMM do not map efficiently onto a modern processor based system. Therefore, an application specific co-processor is normally used to improve performance. This paper outlines two possible FPGA based co-processors for executing core CMM operations based upon a compact bit vector (CBV) data format. This representation significantly increases a system's storage capacity, but reduces processing performance.

A neural supergraph matching architecture is introduced based on relaxation labeling and the minimum common supergraph of pairs of graphs. The system is implemented on correlation matrix memories and is efficient in constructing this supergraph. We test the effectiveness of this graphical cluster representation on two different sets of graphs.

The use of search engines within the Internet is now ubiquitous. This work examines how Grid technology may affect the implementation of search engines by focusing on the Signal Data Explorer application developed within the Distributed Aircraft Maintenance Environment (DAME) project. This application utilizes advanced neural-network-based methods (Advanced Uncertain Reasoning Architecture (AURA) technology) to search for matching patterns in time-series vibration data originating from Rolls-Royce aeroengines (jet engines). The large volume of data associated with the problem required the development of a distributed search engine, where data is held at a number of geographically disparate locations. This work gives a brief overview of the DAME project, the pattern marching problem, and the architecture. It also describes the Signal Data Explorer application and provides an overview of the underlying search engine technology and its use in the aeroengine health-monitoring domain.

In this paper we test a range of three-dimensional face recognition systems, based on the fishersurface method developed in previous work. We show the effect of using a variety of facial surface representations and suggest a method of identifying and extracting useful qualities offered by each system. Combing these components into a unified surface subspace, we create a threedimensional face recognition system producing significantly lower error rates than individual systems tested on the same data. We evaluate systems by performing up to 1,079,715 verification operations on a large test set of 3D face models. Results are presented in the form of false acceptance and false rejection rates, generated by varying a decision threshold applied to a distance metric in combined surface space. 1

We evaluate a new approach to face recognition using a variety of surface representations of three-dimensional facial structure. Applying principal component analysis (PCA), we show that high levels of recognition accuracy can be achieved on a large database of 3D face models, captured under conditions that present typical difficulties to more conventional two-dimensional approaches. Applying a range of image processing techniques we identify the most effective surface representation for use in such application areas as security, surveillance, data compression and archive searching.

In this paper, we propose a simple, flexible, and efficient hybrid spell checking methodology based upon phonetic matching, supervised learning, and associative matching in the AURA neural system. We integrate Hamming Distance and n-gram algorithms that have high recall for typing errors and a phonetic spell-checking algorithm in a single novel architecture. Our approach is suitable for any spell checking application though aimed toward isolated word error correction, particularly spell checking user queries in a search engine. We use a novel scoring scheme to integrate the retrieved words from each spelling approach and calculate an overall score for each matched word. From the overall scores, we can rank the possible matches. We evaluate our approach against several benchmark spellchecking algorithms for recall accuracy. Our proposed hybrid methodology has the highest recall rate of the techniques evaluated. The method has a high recall rate and low-computational cost.

Summary form only given. AURA (Advanced Uncertain Reasoning Architecture) is a generic family of techniques and implementations intended for high-speed approximate search and match operations on large unstructured datasets. AURA technology is fast, economical, and offers unique advantages for finding near-matches not available with other methods. AURA is based upon a high-performance binary neural network called a correlation matrix memory (CMM). Typically, several CMM elements are used in combination to solve soft or fuzzy pattern-matching problems. AURA takes large volumes of data and constructs a special type of compressed index. AURA finds exact and near-matches between indexed records and a given query, where the query itself may have omissions and errors. The degree of nearness required during matching can be varied through thresholding techniques. The PCI-based PRESENCE (Parallel Structured Neural Computing Engine) card is a hardware-accelerator architecture for the core CMM computations needed in AURA-based applications. The card is designed for use in low-cost workstations and incorporates 128 MByte of low-cost DRAM for CMM storage. To investigate the scalability of the distributed AURA system, we implement a word-to-document index of an AURA-based information retrieval system, called MinerTaur, over a distributed PRESENCE CMM.

This paper investigates the practical application of a new method called selfsimilar convolution masks to binary trademark images to produce affine invariant one dimensional histogram descriptions. Because the convolution mask is a scaled version of the original image, the normalised distribution histogram of the resulting grey scale image is an affine invariant description based purely upon the image structure, which can be then be used for database search using a database of binary images. The effectiveness of this approach to trademark image identification is tested. Multiple methods of representing the distribution histograms within the database are also tested for robustness to noise, generalisation and reliability. 1

The Virtual Adaptive Learning Architecture (VALA) project has developed a learning architecture with user interface adaptability that provides a personalized learning environment as well as a repository of reusable learning objects. VALA is designed to be customized easily while allowing for future expansion. Instructors from multiple disciplines can easily find, develop, edit and publish multimedia content in support of their instructional endeavors with VALA. In addition, research questions regarding how the delivery of a learning experience affects learning outcomes can be explored through the system.

We propose a hierarchical clustering algorithm (TreeGCS) based upon the Growing Cell Structure (GCS) neural network of B. Fritzke (1993). Our algorithm refines and builds upon the GCS base, overcoming an inconsistency in the original GCS algorithm, where the network topology is susceptible to the ordering of the input vectors. Our algorithm is unsupervised, flexible, and dynamic and we have imposed no additional parameters on the underlying GCS algorithm. Our ultimate aim is a hierarchical clustering neural network that is both consistent and stable and identifies the innate hierarchical structure present in vector-based data. We demonstrate improved stability of the GCS foundation and evaluate our algorithm against the hierarchy generated by an ascendant hierarchical clustering dendogram. Our approach emulates the hierarchical clustering of the dendogram. It demonstrates the importance of the parameter settings for GCS and how they affect the stability of the clustering.

The structural health of airframes is often monitored by analysis of the frequency of occurrence matrix (FOOM) produced after each flight. Each cell in the matrix records a stress event of a particular severity. These matrices are used to determine how much of the aircraft's life has been used up in each flight. Unfortunately the sensors that produce this data are subject to degradation themselves, resulting in corruption of FOOMs. The paper reports a method of automating detection of sensor faults. It is the only known method that is capable of detecting such faults. The method is in essence a dimensionality reduction algorithm coupled to a novelty detection algorithm that produces measures of unusual counts of stress elements at the level of the individual cell and unusual distributions of counts over the entire FOOM. Cell-level error is detected using a probability threshold and a sum of standard deviations. FOOM-level error is detected using a novel application of the eigenface algorithm. Novelty is measured using a mixture of Gaussian model of the data, fitted using the expectation-maximisation algorithm.

The structural health of airframes is often monitored by analysis of the frequency of occurrence matrix (FOOM) produced after each flight. Each cell in the matrix records a stress event of a particular severity. These matrices are used to determine how much of the aircraft's life has been used up in each flight. Unfortunately, the sensors that produce this data are subject to degradation themselves, resulting in corruption of FOOMs. The paper reports a method of automating detection of sensor faults. It is the only known method that is capable of detecting such faults. The method is in essence a dimensionality reduction algorithm coupled to a novelty detection algorithm that produce measures of unusual counts of stress events at the level of the individual cell and unusual distributions of counts over the entire FOOM. Cell-level error is detected using a probability threshold and a sum of standard deviations. FOOM-level error is detected using a novel application of the eigenface algorithm. Novelty is measured using a Gaussian basis function neural network fitted using the expectation-maximisation algorithm.

Many computational processes require efficient algorithms, those that both store and retrieve data efficiently and rapidly. In this paper we evaluate a selection of data structures for storage efficiency, retrieval speed and partial matching capabilities using a large information retrieval dataset. We evaluate standard data structures, for example inverted file lists and hash tables but also a novel binary neural network that incorporates superimposed coding, associative matching and row-based retrieval. We identify the strengths and weaknesses of the approaches. The novel neural network approach is superior with respect to training speed and partial match retrieval time.

We propose a hierarchical, unsupervised clustering algorithm (TreeGCS) based upon the Growing Cell Structure (GCS) neural network of Fritzke. Our algorithm improves an inconsistency in the GCS algorithm, where the network topology is susceptible to the ordering of the input vectors. We demonstrate improved stability of the GCS foundation by alternating the input vector order on each presentation. We evaluate our automatically produced cluster hierarchy against that generated by an ascendant hierarchical clustering dendogram. We use a small dataset to illustrate how our approach emulates the hierarchical clustering of the dendogram, regardless of the input vector order.

This paper presents an initial investigation of the performance of relaxation by elimination (RBE) using neural networks for matching three dimensional representations of chemical structures (small molecules). The approach differs from traditional schemes in that it uses a binary representation for information, allowing the use of correlation matrix memories (CMMs) for data processing. Furthermore, the use of CMMs allows an inherently parallel system, allowing high-performance processing.

This paper describes a generic and fast classifier that uses a binary CMM (Correlation Matrix Memory) neural network for storing and matching a large amount of patterns efficiently, and a k-NN rule for classification. To meet CMM input requirements, a robust encoding method is proposed to convert numerical inputs into binary ones with the maximally achievable uniformity. To reduce the execution bottleneck, a hardware implementation of the CMM is described, which shows the network with on-board training and testing operates at over 200 times the speed of a current mid-range workstation, and is scaleable to very large problems. The CMM classifier has been tested on several benchmarks and, comparing with a simple k-NN classifier, it gave less than 1 % lower accuracy and over 4 and 12 times speed-ups in software and hardware respectively. 1

This paper describes a statistical framework for the unsupervised learning of linear filter combinations for feature characterisation. The learning strategy is two step. In the first instance, the EM algorithm is used to learn the foreground probability distribution. This is an abductive process, since we have a detailed model of the background process based on the known noise-response characteristics of the filter-bank. The second phase uses the a posteriori foreground and background probabilities to compute a weighted between-class covariance matrix. We use the principal components analysis to locate the linear filter combinations that maximise the between class covariance matrix. The new feature characterisation method is illustrated for the problem of extracting linear features from complex millimetre radar images. The method proves to be effective in learning a mixture of sine and cosine phase Gabor functions necessary to capture shadowed line structures.

This paper describes a novel technique for segmenting frequency modulated line-textures. Textures of this sort abound in nature and are typified by growth patterns in which the deposition rate varies over time. The basic idea underpinning the technique is to use the S-Gabor kernel as a frequency modulated channel response function. According to this channel model, the central frequency changes exponentially with distance from the centre of the kernel. In order to segment the resulting texture response, we use fuzzy clustering to locate peaks in the Fourier power spectrum. In this way we estimate both the centre-frequency and the modulation parameters of the filter bank. We illustrate the effectiveness of our technique on the segmentation of growth patterns on fish scales.

An essential part of image analysis is the identification of objects within the image. This paper describes the results of applying a binary associative memory neural network (ADAM), to the complex task of identifying shapes in document images. The associative memory is used to implement the generalised Hough Transform, exploiting the fast look-up ability of the associative memory to give a high-speed image analysis tool.

We describe a technique for matching a single, learned elastic model of the shape of normal chromosomes to chromosomal images. Our model has a hierarchical organisation, with increasingly coarse shape descriptions at higher levels. A problem of finding the model description most likely to have generated an image is reduced to one of matching the locations of model points to the locations of image features encoded on a Kohonen Self-Organising Map (SOM). During matching, coarse shape information is communicated between levels via fixed, viewpoint-independent transformations between object-based frames. After matching, frame parameters provide a compact, multi-scale description of important shape information. We propose using this information to train a neural network to recognise structurally damaged chromosomes.