Search

This paper considers signal plan selection; the main topic is the design of a system for utilising pattern matching to assist the timely selection of sound signal control plan changes. In this system, historical traffic flow data is continually searched, seeking traffic flow patterns similar to today’s. If, in one of these previous similar situations, (a) the signal plan utilised was different to that being utilised today and (b) it appears that the performance achieved was better than the performance likely to be achieved today, then the system recommends an appropriate signal plan switch. The heart of the system is “similarity”. Two traffic flow patterns (two time series of traffic flows arising from two different days) are said to be “similar” if the distance between them is small; similarity thus depends on how the metric or distance between two time series of traffic flows is defined. A simple example is given which suggests that utilising the standard Euclidean distance between the two sequences comprising cumulatives of traffic flow may be better than utilising the standard Euclidean distance between the original two sequences of traffic flow data. The paper also gives measured on-street public transport benefits which have arisen from using a simple rule-based (traffic-responsive) signal plan selection system, compared with a time-tabled signal plan selection system.

In the study of information flow in the nervous system, component processes can be investigated using a range of electrophysiological and imaging techniques. Although data is difficult and expensive to produce, it is rarely shared and collaboratively exploited. The Code Analysis, Repository and Modelling for e-Neuroscience (CARMEN) project addresses this challenge through the provision of a virtual neuroscience laboratory: an infrastructure for sharing data, tools and services. Central to the CARMEN concept are federated CARMEN nodes, which provide: data and metadata storage, new, thirdparty and legacy services, and tools. In this paper, we describe the CARMEN project as well as the node infrastructure and an associated thick client tool for pattern visualisation and searching, the Signal Data Explorer (SDE). We also discuss new spike detection methods, which are central to the services provided by CARMEN. The SDE is a client application which can be used to explore data in the CARMEN repository, providing data visualization, signal processing and a pattern matching capability. It performs extremely fast pattern matching and can be used to search for complex conditions composed of many different patterns across the large datasets that are typical in neuroinformatics. Searches can also be constrained by specifying text based metadata filters. Spike detection services which use wavelet and morphology techniques are discussed, and have been shown to outperform traditional thresholding and template based systems. A number of different spike detection and sorting techniques will be deployed as services within the CARMEN infrastructure, to allow users to benchmark their performance against a wide range of reference datasets.

In this paper, we show the effect of using a variety of facial surface feature maps within the Fishersurface technique, which uses linear discriminant analysis, and suggest a method of identifying and extracting useful qualities offered by each surface feature map. Combining these multi-feature subspace components into a unified surface subspace, we create a three-dimensional face recognition system producing significantly lower error rates than individual surface feature map 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 surface space.

The availability of high frequency data sets in finance has allowed the use of very data intensive techniques using large data sets in forecasting. An algorithm requiring fast k-NN type search has been implemented using AURA, a binary neural network based upon Correlation Matrix Memories. This work has also constructed probability distribution forecasts, the volume of data allowing this to be done in a nonparametric manner. In assistance to standard statistical error measures the implementation of simulations has allowed actual measures of profit to be calculated.

This paper evaluates a novel k-nearest neighbour (k-NN) classifier built from binary neural networks. The binary neural approach uses robust encoding to map standard ordinal, categorical and numeric data sets onto a binary neural network. The binary neural network uses high speed pattern matching to recall a candidate set of matching records, which are then processed by a conventional k-NN approach to determine the k-best matches. We compare various configurations of the binary approach to a conventional approach for memory overheads, training speed, retrieval speed and retrieval accuracy. We demonstrate the superior performance with respect to speed and memory requirements of the binary approach compared to the standard approach and we pinpoint the optimal configurations.

In this paper, we propose a simple and flexible spell checker using efficient associative matching in the AURA modular neural system. Our approach aims to provide a pre-processor for an information retrieval (IR) system allowing the user's query to be checked against a lexicon and any spelling errors corrected, to prevent wasted searching. IR searching is computationally intensive so much so that if we can prevent futile searches we can minimise computational cost. We evaluate our approach against several commonly used spell checking techniques for memory-use, retrieval speed and recall accuracy. The proposed methodology has low memory use, high speed for word presence checking, reasonable speed for spell checking and a high recall rate.

In this paper, we propose a hierarchical, lexical clustering neural network algorithm that automatically generates a thesaurus (synonym abstraction) using purely stochastic information derived from unstructured text corpora and requiring no prior word classifications. The lexical hierarchy overcomes the Vocabulary Problem by accommodating paraphrasing through using synonym clusters and overcomes Information Overload by focusing search within cohesive clusters. We describe existing word categorisation methodologies, identifying their respective strengths and weaknesses and evaluate our proposed approach against an existing neural approach using a benchmark statistical approach and a human generated thesaurus for comparison. We also evaluate our word context vector generation methodology against two similar approaches to investigate the effect of word vector dimensionality and the effect of the number of words in the context window on the quality of word clusters produced. We demonstrate the effectiveness of our approach and its superiority to existing techniques.

This paper describes a novel search and pattern matching technology, AURA, based on neural networks. The technology exploits the simple and fast training of correlation matrix memories and their ability to match noisy and incomplete data. Unlike other neural network approaches, the method scales well to accommodate very large datasets, and has a simple and effective hardware implementation. It achieves this by storing basic feature data and relying on the fast matching ability of correlation matrix memory methods, as well as a two-stage processing approach. The paper demonstrates that the technology is applicable to a wide range of problems and shows how it has been implemented in dedicated hardware for high-performance applications.

We introduce a theoretical framework for estimating the matching performance of binary correlation matrices acting as hetero-associative memories. The framework is applicable to non-recursive, fully connected systems with binary (0, 1) Hebbian weights and hard-limited threshold. It can handle both full and partial matching of single or multiple data items in non-square memories. Theoretical development takes place under a probability theory framework. Inherent uncertainties in the matching process are accommodated through the use of probability distributions to describe the numbers of correct and incorrect neuron responses during retrieval. Theoretical predictions are verified experimentally for medium-sized memories and are used to aid in the design of larger systems. The results highlight the fact that correlation-based models can act as highly efficient memories provided a small probability of retrieval error is accepted.

We present a technique for initial location of scattered chromosomal objects within multi-resolution images of human blood cells. Kohonen Self Organising Maps learn to extract salient image features in the vicinity of located objects. Feature extraction is to form the first stage in a neural network system applied to the problem of recognizing structural aberrations in chromosomes.

Using probability theory, it is shown that tuple preprocessing in ADAM fails to store patterns in a uniform manner in the first stage memory matrix. Instead, patterns tend to be clustered at the edges of tuple unit addressed matrix regions. A technique for overcoming this is given. Simulations show that this new technique is a great advantage for fault tolerance.

A performance analysis model for regular multicomputer networks is developed which covers both single- or multiple-message accepting PE models and buffered or unbuffered communication switches. The performance of a four-neighbour mesh with ‘wrapped-around’ boundaries as found in most transputer nets is evaluated in terms of the overall bandwidth vs. the message generation rate and ‘span’ radii. We show that the network will not saturate if the message generation rate is less than unity and the message hopcount is less than six.

Two-colour time-resolved resonance Raman spectroscopy has been used to probe the longest-lived excited state of iron(III) tetraphenylporphyrin (Fe(TPP)Cl). Single-laser (408 nm) irradiation of Fe(TPP)Cl showed that significant depopulation of the ground electronic state could be achieved by use of high power, 10 ns duration, pulses. Two-colour time-resolved spectra showed the formation of an excited species whose lifetime was shorter than the laser pulse widths used. The resonance Raman spectra of this species, identified as the lowest-lying “triplet” state, showed frequency shifts associated with porphyrin core expansion. The origins of the differences between the resonance Raman spectra of “triplet” Fe(TPP)Cl and triplet Zn(TPP) are discussed.

A distributed associative memory system which is ideal for scene analysis is described. Recall of associated patterns using incomplete originals is made possible by the use of a distributed storage mechanism and a novel recall procedure. The memory is shown to store associations between patterns more efficiently than a conventional file store. The paper describes the memory structure, the recall process and its storage abilities, as well as an example of its implementation in hardware.