Abstract

Aerial drones are increasingly being considered as a valuable tool for inspection in safety critical contexts. Nowhere is this more true than in mining operations which present a dynamic and dangerous environment for human operators. Drones can be deployed in a number of contexts including efficient surveying as well as search and rescue missions. Operating in these dynamic contexts is challenging however and requires the drones control software to detect and adapt to conditions at run-time. To help in the development of such systems we present Aloft, a simulation supported testbed for investigating self-adaptive controllers for drones in mines. Aloft utilises the Robot Operating system (ROS) and a model environment using Gazebo to provide a physics-based testing. The simulation environment is constructed from a 3D point cloud collected in a physical mock-up of a mine and contains features expected to be found in real-world contexts. Aloft allows members of the research community to deploy their own self-adaptive controllers into the control loop of the drone to evaluate the effectiveness and robustness of controllers in a challenging environment. To demonstrate our system we provide a self-adaptive drone controller and operating scenario as an exemplar. The self-adaptive drone controller provided utilises a two-layered architecture with a MAPE-K feedback loop. The scenario is an inspection task during which we inject a communications failure. The aim of the controller is to detect this loss of communication and autonomously perform a return home behaviour. Limited battery life presents a constraint on the mission, which therefore means that the drone should complete its mission as fast as possible. Humans, however, might also be present within the environment. This poses a safety risk and the drone must be able to avoid collisions during autonomous flight. In this paper we describe the controller framework and the simulation environment and provide information on how a user might construct and evaluate their own controllers in the presence of disruptions at run-time.

"© {C. Imrie et al. | ACM} 2024. This is the author's version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record was published in SEAMS'24, https://doi.org/10.1145/3643915.3644107"

Abstract

Medical staff shortages and growing healthcare demands due to an ageing population mean that many patients face delays in receiving critical care in the emergency departments (EDs) of hospitals worldwide. As such, the use of autonomous, robotics and AI technologies to help streamline the triage of ED patients is of utmost importance. In this paper, we present our ongoing work to develop an autonomous emergency triage support system intended to alleviate the current pressures faced by hospital emergency departments. By employing a combination of robotic and AI techniques, our solution aims to speed up the initial stages of ED triage. Its preliminary evaluation using synthetic patient datasets generated with ED medic input suggests that our solution has the potential to improve the ED triage process, supporting the timely and accurate delivery of patient care in emergency settings..

“© 2023 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.”

Abstract

Esports have become major international sports with hundreds of millions of spectators. Esports games generate massive amounts of telemetry data. Using these to predict the outcome of esports matches has received considerable attention, but micro-predictions, which seek to predict events inside a match, is as yet unknown territory. Micro-predictions are however of perennial interest across esports commentators and audience, because they provide the ability to observe events that might otherwise be missed: esports games are highly complex with fast-moving action where the balance of a game can change in the span of seconds, and where events can happen in multiple areas of the playing field at the same time. Such events can happen rapidly, and it is easy for commentators and viewers alike to miss an event and only observe the following impact of events. In Dota 2, a player hero being killed by the opposing team is a key event of interest to commentators and audience. We present a deep learning network with shared weights which provides accurate death predictions within a five-second window. The network is trained on a vast selection of Dota 2 gameplay features and professional/semi-professional level match dataset. Even though death events are rare within a game (1\% of the data), the model achieves 0.377 precision with 0.725 recall on test data when prompted to predict which of any of the 10 players of either team will die within 5 seconds. An example of the system applied to a Dota 2 match is presented. This model enables real-time micro-predictions of kills in Dota 2, one of the most played esports titles in the world, giving commentators and viewers time to move their attention to these key events.

Abstract

This paper presents an architecture for the creation of emotionally congruent music using machine learning aided sound synthesis. Our system can generate a small corpus of music using Hidden Markov Models; we can label the pieces with emotional tags using data elicited from questionnaires. This produces a corpus of labelled music underpinned by perceptual evaluations. We then analyse participant’s galvanic skin response (GSR) while listening to our generated music pieces and the emotions they describe in a questionnaire conducted after listening. These analyses reveal that there is a direct correlation between the calmness/scariness of a musical piece, the users’ GSR reading and the emotions they describe feeling. From these, we will be able to estimate an emotional state using biofeedback as a control signal for a machine-learning algorithm, which generates new musical structures according to a perceptually informed musical feature similarity model. Our case study suggests various applications including in gaming, automated soundtrack generation, and mindfulness.

Abstract

This research explores and designs an effective experimental interface to evaluate people's emotional responses to horror music. We studied methodological approaches by using traditional psychometric techniques to measure emotional responses, including self-reporting, and galvanic skin response (GSR). GSR correlates with psychological arousal. It can help circumvent a problem in self-reporting where people are unwilling to report particular felt responses, or confuse perceived and felt responses. We also consider the influence of familiarity. Familiarity can induce learned emotional responses rather than listeners describing how it actually makes them feel. The research revealed different findings in self-reports and GSR data. Both measurements had an interaction between music and familiarity but show inconsistent results from the perspective of simple effects.

Abstract

Esports – video games played competitively that are broadcast to large audiences – are a rapidly growing new form of mainstream entertainment. Esports borrow from traditional TV, but are a qualitatively different genre, due to the high flexibility of content capture and availability of detailed gameplay data. Indeed, in esports, there is access to both real–time and historical data about any action taken in the virtual world. This aspect motivates the research presented here, the question asked being: can the information buried deep in such data, unavailable to the human eye, be unlocked and used to improve the live broadcast compilations of the events? In this paper, we present a large–scale case study of a production tool called Echo, which we developed in close collaboration with leading industry stakeholders. Echo uses live and historic match data to detect extraordinary player performances in the popular esport Dota 2, and dynamically translates interesting data points into audience–facing graphics. Echo was deployed at one of the largest yearly Dota 2 tournaments, which was watched by 25 million people. An analysis of 40 hours of video, over 46,000 live chat messages, and feedback of 98 audience members showed that Echo measurably affected the range and quality of storytelling, increased audience engagement, and invoked rich emotional response among viewers.

Abstract

As part of their design, card games often include information that is hidden from opponents and represents a strategic advantage if discovered. A player that can discover this information will be able to alter their strategy based on the nature of that information, and therefore become a more competent opponent. In this paper, we employ association rule-mining techniques for predicting item multisets, and show them to be effective in predicting the content of Netrunner decks. We then apply different modifications based on heuristic knowledge of the Netrunner game, and show the effectiveness of techniques which consider this knowledge during rule generation and prediction.

Abstract

In this paper we introduce an environmentally driven conceptual framework of Business Model change. Business models acquired substantial momentum in academic literature during the past decade. Several studies focused on what exactly constitutes a Business Model (role model, recipe, architecture etc.) triggering a theoretical debate about the Business Model’s components and their corresponding dynamics and relationships. In this paper, we argue that for Business Models as cognitive structures, are highly influenced in terms of relevance by the context of application, which consequently enriches its functionality. As a result, the Business Model can be used either as a role model (benchmarking) or a recipe (strategy). For that purpose, we assume that the Business Model is embedded within the economic (task) environment, and consequently affected by it. Through a typology of the environmental impact on the Business Model productivity, we introduce a conceptual framework that aims to capture the salient features of Business Model emergent resilience as reaction to two types impact: productivity constraining and disturbing.

Abstract

The global video games industry has experienced and exponential growth in terms of socioeconomic impact during the last 50 years. Surprisingly, little academic interest is directed towards the industry, particularly in the context of BM Change. As a technologically intensive creative industry, developing studios and publishers experience substantial internal and external forces to identify, and sustain, their competitive advantage. To achieve that, managers are called to systematically explore and exploit, alternative BMs that are compatible with the company’s strategy. We build on empirical analysis of the video–games industry to construct a Toolkit that i) will help practitioners and academics to describe the industrial ecosystem of BMs more accurately, and ii) use it a strategic roadmap for managers to navigate through alternatives for entrepreneurial and growth purposes.

Abstract

In this paper, we introduce an implementation of the attribute selection algorithm, Correlation-based Feature Selection (CFS) integrated with our k–nearest neighbour (k–NN) framework. Binary neural networks underpin our k–NN and allow us to create a unified framework for attribute selection, prediction and classification. We apply the framework to a real world application of predicting bus journey times from traffic sensor data and show how attribute selection can both speed our k–NN and increase the prediction accuracy by removing noise and redundant attributes from the data.

Abstract

This paper describes recent enhancements to the YouShare platform, the online collaboration environment, which allows researchers to share data and software applications and perform compute–intensive analysis tasks quickly and securely. The enhancements to the platform are a result of user feedback on the current system and technology advancements. These fall into four groups – better handling of searching, use of synonyms, the addition of a workflow tool and enhancements to the infrastructure. The paper outlines these improvements.

Abstract

The aim of our work is to develop Intelligent Decision Support (IDS) tools and techniques to convert traffic data into intelligence to assist network managers, operators and to aid the travelling public. The IDS system detects traffic problems, identifies the likely cause and recommends suitable interventions which are most likely to mitigate congestion of that traffic problem. In this paper, we propose to extend the existing tools to include dynamic hierarchical and distributed processing; algorithm optimisation using natural computation techniques; and, using a meta–learner to short–circuit the optimisation by learning the best settings for specific data set characteristics and using these settings to initialise the GA.

Abstract

Pattern recognition often relies on the distance between patterns, and the Euclidean distance is frequently used. In traffic studies, the sequence of flow cumulatives is more directly related to the character of the flow (e.g., congested or free), and might be expected to supply more meaningful traffic information than would a sequence of unsummed flows. However, if there is noise in the detection procedure, an error in the first count of a sequence would be carried forward to subsequent cumulatives, and would therefore, like errors in general, have a greater effect on the Euclidean distance between cumulatives than it would on that between the raw flows. This paper aims at providing a quantitative measure of the classification errors caused by noise on the Euclidean distance between sequences of cumulatives compared to those arising from distance calculations between raw sequences, and thus supplying a guideline for the situation when cumulatives or raw flows are to used in the presence of noise.

Abstract

This paper considers models of within day and day-to-day driver (or traveller) decisions; and focusses on a splitting rate method of modelling these decisions sequentially. The paper begins with a dynamical system proposed in Smith (1984), using route-swapping, and shows how this same system may be applied within a splitting rate model. Some stability and convergence results are to be presented. Finally the paper suggests how signal control may be introduced in a simple and helpful way. This combined routeing / signal control model may be utilized quite easily to yield initial “suggested” signal control interventions corresponding to specific incidents.

Abstract

This paper presents a binary neural network algorithm for short-term traffic flow prediction. The algorithm can process both univariate and multivariate data from a single traffic sensor using time series prediction (temporal lags) and can combine information from multiple traffic sensors with time series prediction (spatial–temporal lags). The algorithm provides Intelligent Decision Support (IDS) for road network managers to proactively manage problems on the network as the predictions generated may be used to determine if traffic control interventions need to be applied. The algorithm can operate in near-real-time and dynamically; using data from UTC or UTMC systems. It is based on the Advanced Uncertain Reasoning Architecture (AURA) k–nearest neighbour prediction algorithm, which is designed for scalability and fast performance. The AURA k–NN predictor outperforms other machine learning techniques with respect to prediction accuracy and is able to train and predict rapidly. The basic AURA k–NN time series prediction algorithm was extended by incorporating average daily profiles and variable weighting into the prediction in this paper. The average daily profile of a variable is calculated as the average reading of the variable for a particular time of day and day of the week after removing outliers. When data vectors are matched in the AURA k–NN, the daily profile adds an extra dimension to the match. This process was further enhanced by weighting the profile using variable weighting to vary the profile’s significance. It is shown that incorporating these two additional aspects improves the accuracy of the prediction compared to the standard AURA k–NN, resulting in a very fast and accurate traffic prediction tool.

Abstract

Urban traffic control systems such as the widely deployed SCOOT system, incrementally respond to changing traffic conditions. Such systems are often complemented by traffic control centres where road network managers intervene manually to mitigate rapidly developing congestion events. An Intelligent Decision Support (IDS) system developed by the authors within the UK FREEFLOW (FF) project to aid the network managers is presented in this paper. The primary objective of the FF IDS system is to identify traffic congestion in near-real-time and to recommend appropriate traffic control intervention measures. The FF–IDS consists of multiple internal components. A state estimation component monitors live traffic sensor data and determines if there is a congestion problem on the road network. If a problem is identified, a binary neural pattern-matching component is used to identify past time periods with similar congestion events. This is able to rapidly search large historic traffic datasets finding sets of traffic control interventions carried out during similar historical time periods. The effectiveness of each intervention is evaluated using a Performance Index (PI) and the intervention that resulted in the highest improvement in PI is recommended to network managers. The FF-IDS system can also present traffic incidents and equipment faults that occurred during these historical time periods to the network manager as potential causes of the problem. This paper describes the FF–IDS system in detail. The system is currently under development. An early version of the FF–IDS system was trialled using off-line data from London, yielding encouraging preliminary results.

Abstract

Many Condition Monitoring (CM) domains are suffering from the dual challenges of substantial increases in the volumes of data being produced and collected by sensing systems, and the challenges of modelling increasing complexity in the remote monitored systems. These two issues give rise to the problem that fast and reliable data mining of CM data is a computationally demanding task for real–time (or near real–time) applications. We present the use of AURA [1], a class of binary associative network built on correlation matrix memories (CMMs), as an underpinning technology for efficient, scalable pattern recognition in complex and large scale CM applications. AURA is a class of binary neural network. However, it has a number of advantages over standard neural network techniques for CM pattern classification tasks. These include; high levels of data compression, one-pass training for on–line training, a scalable architecture that can be readily mapped onto high performance computing platforms, and a sound theoretical basis to determine the bounds of the system operation. We describe applications illustrating how the AURA system can be optimised to create an extremely efficient and scalable k–nearest neighbour classifier for multi–variate models. We will also illustrate how the one-pass training capability of the AURA system can be used as the basis of normality and exception modelling in complex CM systems. This latter application has particularly powerful advantages for fault detection models in domains which are characterised by highly dynamic trends or drifting in the standard operational mode of a system, and which, as a result, are extremely difficult to accurately model. The application of the AURA techniques will be illustrated with industry led exemplars in the transport and energy sectors.

Abstract

Despite the vast wealth of traffic data available, currently there is only limited integration, analysis and utilisation of data in the transport domain. Yet, accurate congestion and incident detection is vital for traffic network operators to allow them to mitigate the cost of traffic incidents. Recurrent (cyclical) traffic congestion tends to be managed using timetabled control measures or through the use of adaptive traffic control systems such as SCOOT and SCATS. However, for non-recurrent congestion with rapid onset, such as the congestion caused by a traffic incident or traffic equipment failure, traffic network operators have to quickly detect the problem and then determine the likely cause before selecting the most appropriate action to both manage the traffic network and mitigate the congestion. This is a complex task requiring specialist knowledge where assistance from automated tools will help facilitate the operator tasks. Automated detection is becoming an increasingly viable option due to the increased use of traffic sensors in the road network. Therefore, the aim of the FREEFLOW project is to provide an Intelligent Decision Support (IDS) tool which is designed to complement existing fixed-time traffic control systems and adaptive systems SCOOT and SCATS. IDS will use traffic sensor data to rapidly identify traffic problems, recommend appropriate interventions that worked in the past for similar problems and assist the traffic network operators to pinpoint the cause of the problem. Recommendations will be displayed to the network operator who will use this knowledge to select the most appropriate course of action. This paper describes and analyses the components of the IDS tool used for identifying incidents and faulty equipment.

Abstract

The k-nearest neighbor algorithm (k–NN) has been used in the literature for traffic state estimation and prediction over the last decade or so. A number of such multivariate methods use input data from more than one traffic sensor. While a significant amount of discussion can be found in the literature aiming towards optimising the parameters of the k–NN for better accuracy of such models, limited research is available on configuring the k–NN to differentiate between different spatial patterns in the multivariate models. This paper presents an approach to distinguish spatial patterns from one another reliably in traffic variables observed using a number of point–based sensors in a neighbourhood of road links. The application of the proposed approach is demonstrated using AURA, a fast binary pattern matching tool based on neural networks. Two different spatial patterns of traffic congestion plus non–congested situations are simulated using a PARAMICS micro–simulation model. The AURA software is used to identify similar time periods of congestion using data from a congested time period as input using conventional and proposed distance metrics. It is shown that the proposed distance metrics can identify different spatial congestion patterns better than conventional methods. This method will be useful for traffic estimation and prediction methods that use the k–nearest neighbor algorithm or its variants.

Abstract

Adaptive traffic control systems such as SCOOT and SCATS are designed to respond to changes in traffic conditions and provide heuristically optimised traffic signal settings. However, these systems make gradual changes to signal settings in response to changing traffic conditions. In the EPSRC and TSB funded FREEFLOW project, a tool is being designed to rapidly identify severe traffic problems using traffic sensor data and recommend traffic signal plans and UTC parameters that have worked well in the past under similar traffic conditions for immediate implementation. This paper will present an overview of this tool, called the Intelligent Decision Support (IDS),that is designed to complement adaptive traffic control systems. The IDS is essentially a learning based system. It requires an historic database of traffic sensor data and traffic control intervention data for the application area as a knowledge base. The IDS, when deployed online, will monitor traffic sensor data to determine if the network is congested using traffic state estimation models. When IDS identifies congestion in the network, the historic database is queried for similar congestion events, where the similarity is based on both the severity and the spatial pattern of congestion. Traffic control interventions implemented during similar congestion events in the historic database are then evaluated for their effectiveness to mitigate congestion. The most effective traffic control interventions are recommended by IDS for implementation, along with an associated confidence indicator. The IDS is designed to work online against large historic datasets, and is based on traffic state estimation models developed at Imperial College London and pattern matching tools developed at the University of York. The IDS is tested offline using Inductive Loop Detector (ILD) data obtained from the ASTRID system and traffic control intervention data obtained from the UTC system at Transport for London (TfL) during its development. This paper presents the preliminary results using TfL data and outlines future research avenues in the development of IDS.

Abstract

Although a substantial amount of research has examined the constructs of warmth and competence, far less has examined how these constructs develop and what benefits may accrue when warmth and competence are cultivated. Yet there are positive consequences, both emotional and behavioral, that are likely to occur when brands hold perceptions of both. In this paper, we shed light on when and how warmth and competence are jointly promoted in brands, and why these reputations matter.

Abstract

This report describes preliminary analysis of strategies to activate and deactivate a bus pre-signal using vehicle count data. The bus pre–signal currently operates during preset times to regulate access to a length of road controlled at the other end by vehicle-actuated traffic signals. However, vehicle flows at the pre–signal vary on a daily basis so a more demand-based approach would be more effective. There has been much research performed to optimise pre–signal cycle times and bus priority at pre–signals. We focus on identifying the optimal strategy to activate and deactivate the bus pre–signal using vehicle demand rather than the current fixed time strategy. The ideal strategy should be stable, robust, consistent and timely. We investigate strategies using vehicle counts, queueing theory and estimation and prediction. Our recommended strategy combines aspects of all three areas.

Abstract

The twin problems of congestion and inner-city parking limitations affect many cities. One solution is to promote the use of Park and Ride sites. However, for effective use of the sites, drivers need to know where the sites are and which is the "best" site to use. This work introduces a methodology to pinpoint and guide drivers to the best Park and Ride site from their current location. While drivers may be able to obtain traffic, car park location and free space data individually, the information is not usually coordinated. By fusing up–to–date details of traffic jams, roadworks and accidents coupled with free parking spaces and combining this with a novel route weighting methodology, we are able to ensure that intelligent information is displayed to guide drivers. The method uses optimised data structures and proprietary scoring measures to ensure it is fast and accurate. The method provides a simple and low cost solution through the use of existing technologies to display information to drivers.

Abstract

In this paper we focus on identifying image structures at different levels in figurative (trademark) images to allow higher level similarity between images to be inferred. To identify image structures at different levels, it is desirable to be able to achieve multiple views of an image at different scales and then extract perceptually–relevant shapes from the different views. The three aims of this work are: to generate multiple views of each image in a principled manner, to identify structures and shapes at different levels within images and to emulate the Gestalt principles to guide shape finding. The proposed integrated approach is able to meet all three aims.

Abstract

In this paper, we develop a system to classify the outputs of image segmentation algorithms as perceptually relevant or perceptually irrelevant with respect to human perception. The work is aimed at figurative images. We previously investigated human visual perception of trademark images and established a body of ground truth data in the form of trademark images and their respective human segmentations. The work indicated that there is a core set of segmentations for each image that people perceive. Here we use this core set of segmentations to train a classifier to classify closed shapes output from an image segmentation algorithm so that the method returns the image segments that match those produced by people. We demonstrate that a perceptual relevance classifier is attainable and identify a good methodology to achieve this. The paper compares MLP, SVM, Bayes and regression classifiers for classifying shapes. MLPs perform best with an overall accuracy of 96.4%.

Abstract

Ensuring the uniqueness of trademark images and protecting their identities are the most important objectives for the trademark registration process. To prevent trademark infringement, each new trademark must be compared to a database of existing trademarks. Given a newly designed trademark image, trademark retrieval systems are not only concerned with finding images with similar shapes but also locating images with similar layouts. Performing a linearsearch, i.e., computing the similarity between the query and each database entry and selecting the closest one, is ineffi- cient for large database systems. An effective and efficient indexing mechanism is, therefore, essential to select a small collection of candidates. This paper proposes a framework in which a graph-based indexing schema will be applied to facilitate efficient trademark retrieval based on spatial relations between image components, regardless of mutual shape similarity. Our framework starts by segmenting trademark images into distinct shapes using a shape identification algorithm. Identified shapes are then encoded automatically into an attributed graph whose vertices represent shapes and whose edges show spatial relations (both directional and topological) between the shapes. Using a graph–based indexing schema, the topological structure of the graph as well as that of its subgraphs are represented as vectors in which the components correspond to the sorted Laplacian eigenvalues of the graph or subgraphs. Having established the signatures, the indexing amounts to a nearest neighbour search in a model database. For a query graph and a large graph data set, the indexing problem is reformulated as that of fast selection of candidate graphs whose signatures are close to the query signature in the vector space. An extensive set of recognition trials, including a comparison with manually constructed graphs, show the efficacy of both the automatic graph construction process and the indexing schema.

Abstract

In this paper, we introduce a neural network-based shape matching algorithm that uses Johnson Counter codes coupled with chain codes. Shape matching is a fundamental requirement in content-based image retrieval systems. Chain codes describe shapes using sequences of numbers. They are simple and flexible. We couple this power with the efficiency and flexibility of a binary associative–memory neural network. We focus on the implementation details of the algorithm when it is constructed using the neural network. We demonstrate how the binary associative–memory neural network can index and match chain codes where the chain code elements are represented by Johnson codes.

Abstract

In this paper, we investigate human visual perception and establish a body of ground truth data elicited from human visual studies. We aim to build on the formative work of Ren, Eakins and Briggs who produced an initial ground truth database. Human participants were asked to draw and rank their perceptions of the parts of a series of figurative images. These rankings were then used to score the perceptions, identify the preferred human breakdowns and thus allow us to induce perceptual rules for human decomposition of figurative images. The results suggest that the human breakdowns follow well–known perceptual principles in particular the Gestalt laws.

Abstract

In this paper, we introduce a neural network-based decision table algorithm. We focus on the implementation details of the decision table algorithm when it is constructed using the neural network. Decision tables are simple supervised classifiers which, Kohavi demonstrated, can outperform state–of–the–art classifiers such as C4.5. We couple this power with the efficiency and flexibility of a binary associative–memory neural network. We demonstrate how the binary associative-memory neural network can form the decision table index to map between attribute values and data records. We also show how two attribute selection algorithms, which may be used to pre–select the attributes for the decision table, can easily be implemented within the binary associative–memory neural framework. The first attribute selector uses mutual information between attributes and classes to select the attributes that classify best. The second attribute selector uses a probabilistic approach to evaluate randomly selected attribute subsets.

Abstract

The k-Nearest Neighbour (kNN) approach is a widely-used technique for pattern classification. Ranked distance measurements to a known sample set determine the classification of unknown samples. Though effective, kNN, like most classification methods does not scale well with increased sample size. This is due to their being a relationship between the unknown query and every other sample in the data space. In order to make this operation scalable, we apply AURA to the kNN problem. AURA is a highly–scalable associative-memory based binary neural-network intended for high-speed approximate search and match operations on large unstructured datasets. Previous work has seen AURA methods applied to this problem as a scalable, but approximate kNN classifier. This paper continues this work by using AURA in conjunction with kernel-based input vectors, in order to create a fast scalable kNN classifier, whilst improving recall accuracy to levels similar to standard kNN implementations.

Abstract

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.

Abstract

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. This paper continues the AURA II (Advanced Uncertain Reasoning Architecture) project’s research into distributed binary Correlation Matrix Memory (CMM) based upon the PRESENCE (PaRallEl Structured Neural Computing Engine) hardware architecture [14]. Previous work has described how CMMs can be seamlessly implemented onto multiple hardware PRESENCE cards to accelerate core CMM operations. To demonstrate the system, this paper describes how a novel CMM-based information retrieval (IR) system, called MinerTaur, was implemented using multiple PRESENCE cards distributed across a cluster.

Abstract

n 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 evaluate our approach against several benchmark spell-checking algorithms for recall accuracy. Our proposed hybrid methodology has the joint highest top 10 recall rate of the techniques evaluated. The method has a high recall rate and low computational cost.

Abstract

Over the years the amount and range of electronic text stored on the WWW has expanded rapidly, overwhelming both users and tools designed to index and search the information. It is impossible to index the WWW dynamically at query time due to the sheer volume so the index must be pre–compiled and stored in a compact but incremental data structure as the information is ever–changing. Much of the text is unstructured so a data structure must be constructed from such text, storing associations between words and the documents that contain them. The index must be able to index fine–grained word-based associations and also handle more abstract concepts such as synonym groups. A search tool is also required to link to the index and enable the user to pinpoint their required information. We describe such a system we have developed in an integrated hybrid neural architecture and evaluate our system against the benchmark SMART system for retrieval accuracy: recall and precision.

Abstract

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 dendrogram. We use a small dataset to illustrate how our approach emulates the hierarchical clustering of the dendrogram, regardless of the input vector order.

Abstract

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.