• Refine Query
  • Source
  • Publication year
  • to
  • Language
  • 76
  • 69
  • 27
  • 3
  • 1
  • Tagged with
  • 1575
  • 256
  • 191
  • 127
  • 122
  • 115
  • 96
  • 94
  • 90
  • 79
  • 71
  • 60
  • 60
  • 59
  • 50
  • About
  • The Global ETD Search service is a free service for researchers to find electronic theses and dissertations. This service is provided by the Networked Digital Library of Theses and Dissertations.
    Our metadata is collected from universities around the world. If you manage a university/consortium/country archive and want to be added, details can be found on the NDLTD website.
21

Dictionary learning for signal classification

Zubair, Syed January 2014 (has links)
Signal classification is widely applied in science and engineering such as in audio and visual signal processing. The performance of a typical classification system depends highly on the features (used to represent a signal in a lower dimensional space) and the classification algorithms (used to determine the category of the signal based on the features). Recent developments show that dictionary learning based sparse representation techniques have the potential to offer improved performance over the conventional techniques for feature extraction, such as mel frequency cepstrum coefficient (MFCC) and classifier design, such as support vector machine (SVM). In this thesis, we focus on dictionary learning based methods for signal classification and address several challenges as explained below. First we study the potential of using dictionary learning algorithms such as K-SVD for sparse feature extraction obtained by Orthogonal Matching Pursuit (OMP). Specifically, we have proposed the use of pooling and sampling techniques in audio domain to unify the dimension of feature vectors, and to improve computational efficiency. The proposed algorithm is also shown to have advantages for noisy signal classification. Most dictionary learning algorithms have been developed for vector/matrix form of data. Our second contribution is to extend dictionary learning algorithms for high dimensional tensor data and use them to design classifiers. Different from existing tensor dictionary learning methods, we introduce various constraints on the dictionary learning process such as structured sparsity constraints on the core tensor and discriminative constraints on the dictionaries based on the data-spread information measured by Fisher criterion. Such constraints facilitate the design of discriminative classifiers based on reconstruction error and further improve the overall performance even with reduced amount of training data. Recently, structured block sparsity in vector/matrix based dictionary learning method has been shown to outperform signal classification in terms of non-block sparse reconstruction error. In our third contribution, we extend the concept of structured-block sparsity to tensors by providing underlying dictionaries with block structure. We develop an algorithm for structured block-sparse tensor representation and perform classification based upon the block sparse tensor reconstruction error. Our algorithm shows improved performance over its matrix based counter-parts and comparable performance with our previous tensor based method. Our dictionary learning based classification methods are applied on audio and image data for various application scenarios such as speech and music discrimination, speaker identification, digit and face recognition. The experimental results confirm the advantage of the proposed algorithms over several state-of-the-art baseline algorithms.
22

Extending Cartesian genetic programming : multi-expression genomes and applications in image processing and classification

Cattani, Philip Thomas January 2014 (has links)
Genetic Programming (GP) is an Evolutionary Computation technique. Genetic Programming refers to a programming strategy where an artificial population of individuals represent solutions to a problem in the form of programs, and where an iterative process of selection and reproduction is used in order to evolve increasingly better solutions. This strategy is inspired by Charles Darwin's theory of evolution through the mechanism of natural selection. Genetic Programming makes use of computational procedures analogous to some of the same biological processes which occur in natural evolution, namely, crossover, mutation, selection, and reproduction. Cartesian Genetic Programming (CGP) is a form of Genetic Programming that uses directed graphs to represent programs. It is called 'Cartesian', because this representation uses a grid of nodes that are addressed using a Cartesian co-ordinate system. This stands in contrast to GP systems which typically use a tree-based system to represent programs. In this thesis, we will show how it is possible to enhance and extend Cartesian Genetic Programming in two ways. Firstly, we show how CGP can be made to evolve programs which make use of image manipulation functions in order to create image manipulation programs. These programs can then be applied to image classification tasks as well as other image manipulation tasks such as segmentation, the creation of image filters, and transforming an input image in to a target image. Secondly, we show how the efficiency - the time it takes to solve a problem - of a CGP program can sometimes be increased by reinterpreting the semantics of a CGP genome string. We do this by applying Multi-Expression Programming to CGP.
23

Discovering patterns and anomalies in graphs with discrete and numeric attributes

Davis, Michael January 2014 (has links)
In this thesis, we investigate pattern mining and anomaly detection in datasets with both structural and numeric attributes. Graphs are used to represent complex structures such as social networks, infrastructure networks, information networks and chemical compounds. Many graph datasets are annotated with numeric labels or weights. We show that numeric attributes are closely related to graph structure, and exploit this observation for substructure discovery and anomaly detection. Our first contribution is Agwan (Attribute Graphs: Weighted and Numeric), a generative model for random graphs with discrete labels and weighted edges. We present algorithms for parameter fitting and graph generation. Using real-world directed and undirected graphs as input, we compare our approach to state-of-the-art random graph generators and draw conclusions about the contribution of vertex labels and edge weights to graph structure. Our second contribution is a constraint on substructure discovery based on the 'outlierness" of graph numeric attributes, which is used to improve search performance and discrimination. In our experiments, we implement our method as a pre-processing step to prune anomalous vertices and edges prior to graph mining, allowing us to evaluate it on graph databases and Single Large Graphs. We measure the effect on runtime, memory requirements and coverage of discovered patterns, relative to the unconstrained approaches. We also outline how our method can be extended for high-dimensional numeric features. Finally, we present Yagada (Yet Another Graph-based Anomaly Detection Algorithm), an algorithm to search for anomalies in graphs with numeric attributes. Yagada is explained using several security-related examples and validated with experiments on a physical Access Control database. Quantitative analysis shows that in the upper range of anomaly thresholds, Yagada detects twice as many anomalies as the best-performing numeric discretization algorithm. Qualitative evaluation shows that the detected anomalies are meaningful, representing a combination of structural irregularities and numerical outliers.
24

Egocentric activity recognition using gaze

Hipiny, Irwandi January 2013 (has links)
When coupled with an egocentric camera, a gaze tracker provides the image point of where the person is fixating at. While performing a familiar task, we tend to fixate on activity- relevant objects at the points in time required in the task at hand. The resulting sequence of gaze regions is therefore very useful for inferring the subject 's activity and action class. This thesis addresses the problem of visual recognition of human activity and action from an egocentric point of view. The higher level task of activity recognition is based on processing the entire sequence of gaze regions as users perform tasks such as cooking or assembling objects, while the mid-level task of action recognition , such as pouring into a cup, is addressed via the automatic segmentation of mutually exclusive sequences prior to recognition. Temporal segmentation is performed by tracking two motion based features inside the successive gaze regions. These features model the underlying structure of image motion data at natural temporal cuts. This segmentation is further improved by the incorporation of a 2D color histogram based detection of human hands inside gaze regions . The proposed method learns activity and action models from the sequence of gaze regions. Activities are learned as a bag of visual words, however we introduce a multi-voting scheme to reduce the effect of noisy matching. Actions are, in addition, modeled as a string of visual words which enforces the structural constraint of an action. We introduce contextual information in the form of location based priors. Furthermore, this thesis addresses the problem of measuring task performance from gaze region modeling. The hypothesis is that subjects with greater task performance scores demonstrate specific gaze patterns as they conduct the task, which is expected to indicate the presence of domain knowledge. This may be reflected in for example requiring minimal visual feedback during the completion of a task. This consistent and strategic use of gaze produces nearly identical activity models among those that score higher, whilst a greater variation is observed between models learned from subjects that have performed less well in the given task. Results are shown on datasets captured using an egocentric gaze tracker with two cameras, a frontal facing camera that captures the scene, and an inward facing camera that tracks the movement of the pupil to estimate the subject's gaze fixation. Our activity and action recognition results are comparable to current literature in egocentric activity recognition, and to the best of our knowledge, the results from the task performance evaluation are the first steps towards automatically modeling user performance from gaze patterns.
25

Learning to integrate data from different sources and tasks

Argyriou, A. January 2008 (has links)
Supervised learning aims at developing models with good generalization properties using input/output empirical data. Methods which use linear functions and especially kernel methods, such as ridge regression, support vector machines and logistic regression, have been extensively applied for this purpose. The first question we study deals with selecting kernels appropriate for a specific supervised task. To this end we formulate a methodology for learning combinations of prescribed basic kernels, which can be applied to a variety of kernel methods. Unlike previous approaches, it can address cases in which the set of basic kernels is infinite and even uncountable, like the set of all Gaussian kernels. We also propose an algorithm which is conceptually simple and is based on existing kernel methods. Secondly, we address the problem of learning common feature representations across multiple tasks. It has been empirically and theoretically shown that, when different tasks are related, it is possible to exploit task relatedness in order to improve prediction on each task---as opposed to treating each task in isolation. We propose a framework which is based on learning a common set of features jointly for a number of tasks. This framework favors sparse solutions, in the sense that only a small number of features are involved. We show that the problem can be reformulated as a convex one and can be solved with a conceptually simple alternating algorithm, which is guaranteed to converge to an optimal solution. Moreover, the formulation and algorithm we propose can be phrased in terms of kernels and hence can incorporate nonlinear feature maps. Finally, we connect the two main questions explored in this thesis by demonstrating the analogy between learning combinations of kernels and learning common feature representations across multiple tasks.
26

Hardware neural systems for applications : a pulsed analog approach

Jackson, Geoffrey Bruce January 1996 (has links)
In recent years the considerable interest in the biologically inspired computational paradigm of artificial neural networks has led to a drive to realise these structures as VLSI hardware. At Edinburgh University research has focused on pulse stream neural networks in which neural states are encoded in the time domain as a stream of digital pulses. To date research centred on developing analog CMOS circuits to implement neural functions. In this thesis these circuits are developed and higher, system level issues addressed in order to produce a neural network system suited to use in real-world applications. To discover the key requirements for use in real-world applications, examples of application based hardware systems are reviewed, as is the field of pulse stream neural networks. These requirements led to the design of a VLSI chip, EPSILON II; a pulse stream neural chip optimised for use on the boundary of the analog domain of the real-world and the digital domain of conventional computing. The EPSILON processor card (EPC) places this chip in a system level framework that oversees chip operation and provides the interfaces to analog signals, a standard digital bus and other EPCs. The system level approach taken provides a versatile platform for prototyping applications while operating with minimal host supervision. To demonstrate the versatility of this approach several applications were developed that utilised this hardware. Foremost amongst these was an autonomous mobile robot that utilises the analog nature of the hardware to provide a direct interface to real-world sensors. Also presented are a series of experiments investigating back-propagation learning on a variety of MLP problems. This study reveals the limits and practicalities of training hardware neural networks, in particular the effects of limited weight dynamic range were found to be of primary importance. From this work conclusions are drawn as to the effectiveness and future development of hardware neural computation; specifically the ability to interface to the analog domain and the issues involved in interfacing to conventional computing devices are highlighted.
27

Process-tolerant VLSI neural networks for applications in optimisation

Baxter, Donald J. January 1993 (has links)
Optimisation problems such as scheduling and resource allocation are hard, as large numbers of solutions exist for 'real' problems. Neural networks have been reported to find optimal solutions quickly. These networks derive their power from a massively parallel architecture, drawing its inspiration from the biological nervous system. There is a need for dedicated hardware implementations to accelerate neural computations. There is also a desire to develop autonomous neural systems. Digital circuits are tolerant of process variations. Digital circuits are however large. Analogue circuits are more compact and consume less power, but are dependent on the fabrication process for correct functionality. The choice between the two techniques is determined ultimately by the application. Analogue techniques are necessary to obtain a completely parallel implementation. Both the Hopfield/Tank and Kohonen networks used in optimisation rely upon matched circuit elements. Thus the development of process invariant analogue circuits for neural networks was the major aim of this thesis. Simulations are reported, of the Hopfield/Tank and the Kohonen networks, applied to the 10-city Travelling Salesman Problem (TSP). They confirm the reported optimisation abilities. The Kohonen network is shown to be faster and more robust than the Hopfield/Tank network. The results obtained from two fabricated devices are reported: a small scale test-chip and a large scale, generic building block device (EPSILON). These results show that the circuits developed in this thesis offer a significant immunity to the effects of process variations. The Kohonen network for the TSP was implemented on EPSILON. The Kohonen network was a very tough test for EPSILON, in that it requires a high degree of accuracy to be able to discriminate between the responses of neurons. Process variations prevented EPSILON from solving TSPs greater than 9 cities. The main conclusion of this thesis is that unless the neural algorithm actively compensates for the effects of process variations, the performance of a network implemented in an analogue VLSI is compromised.
28

Asking intelligent questions : the statistical mechanics of query learning

Sollich, P. January 1995 (has links)
This thesis analyses the capabilities and limitations of query learning by using the tools of statistical mechanics to study learning in feed-forward neural networks. In supervised learning, one of the central questions is the issue of generalization: Given a set of training examples in the form of input-output pairs generated by an unknown <I>teacher</I> rule, how can one generate a <I>student</I> which <I>generalizes</I>, i.e., which correctly predicts the outputs corresponding to inputs not contained in the training set? The traditional paradigm has been to study learning from <I>random</I> <I>examples</I>, where training inputs are sampled randomly from some given distribution. However, random examples contain redundant information, and generalization performance can thus be improved by <I>query learning</I>, where training inputs are chosen such that each new training example will be maximally "useful" as measured by a given <I>objective function</I>. We examine two common kinds of queries, chosen to optimize the objective functions, generalization error and entropy (or information), respectively. Within an extended Bayesian framework, we use the techniques of statistical mechanics to analyse the average case generalization performance achieved by such queries in a range of learning scenarios, in which the functional forms of student and teacher are inspired by models of neural networks. In particular, we study how the efficacy of query learning depends on the form of teacher and student, on the training algorithm used to generate students, and on the objective function used to select queries.
29

Ameliorating integrated sensor drift and imperfections : an adaptive 'neural' approach

Tang, Tong Boon January 2006 (has links)
This thesis examines the suggestion that local pre-processing and early classification of high-dimensional sensory signals can be achieved effectively by an artificial neural network (ANN). The multisensor microsystem for a project named “Integrated Diagnostics for Environmental & Analytical Systems (IDEAS)” is used as an example for this study. Four types of electrochemical sensors are implemented and calibrated. In our testbench experiments, the sensory signals are found to experience some stochastic randomness and drift during operation. Therefore, the ANN must be adaptive to allow auto-calibration of the sensors. This thesis reviews existing ANN algorithms to fuse sensory signals and identifies hardware-amenable neural algorithms. The Continuous Restricted Boltzmann machine (CRBM) is chosen as a suitable candidate. The CRBM is further developed in this thesis to facilitate online learning without experiencing <i>Catastrophic Interference (CI) - </i>a known problem in associative memory based models. The CRBM is examined in two separate simulations. The first simulation evaluates the modelling capability of the CRBM while the second simulation focuses on the adaptation of the CRBM to sensor drift in a dynamic environment. The results suggest that the CRBM is able to model high-dimensional, non-Gaussian data distributions with overlapping areas. The CRBM is also compared favourably, in terms of robustness against sensor drift, with trained but subsequently non-adaptive neural models. The thesis also investigates the optimal architecture size and learning rate for the CRBM. Finally, the limitations of the CRBM are studied. The learning rate is identified as the key factor in determining the feasibility of CRBM tracking sensor drift in a dynamic environment.
30

An analogue VLSI study of temporally-asymmetric Hebbian learning

Bofill Petit, Adria January 2005 (has links)
The primary aim of this thesis is to examine whether temporally asymmetric Hebbian learning is analogue VLSI can support temporal correlation learning and spike-synchrony processing. Novel circuits for synapses with spike-timing-dependent plasticity (STDP) are proposed. Results from several learning experiments conducted with a chip containing a small feed-forward network of neurons with STDP synapses are presented. The learning circuits proposed in this thesis can be used to implement weight-independent STDP and learning rules with weight-dependent potentiation. Test results show that the learning windows implemented are very similar to those found in biological neurons. The peaks of potentiation and depression, as well as the decay of both sides of the STDP learning window, can be tuned independently. Therefore, the circuits proposed can be used to explore learning rules with different characteristics. The main challenge for on-chip learning is the long term storage of analogue weights. Previous investigations of temporally asymmetric Hebbian learning rules have shown that weight-independent STDP creates bimodal weight distributions. This thesis investigates the suggestion that the bimodality of the learning rule may render the long-term storage of analogue values unnecessary.  Several experiments have been carried out to study the weight distribution created on-chop. With both weight-independent and moderate weight dependent learning rules the on-chip synapses develop either maximum or zero weights. The results presented show that, in agreement with theoretical analysis of STDP, the mean of the input weight vector decreases with the mean rate of the input spike trains. Some experiments reported indicated that the instability of weight-independent STDP could be used in some applications to maintain the binary weights learnt when the temporal correlations are removed from the inputs. Test results given show that both zero-delay correlations and narrow time windows of correlation can be detected with the hardware neurons. An on-chip two-layer network has been used to detect a hierarchical pattern of temporal correlations embedded in noisy spike trains. The analysis of the activity generated by the network shows that the bimodal weight distribution emerging from STDP learning amplifies the spike synchrony of the inputs.

Page generated in 0.0337 seconds