In this thesis an eye-directed controller is developed that slaves the narrow field display within a split-foveal-peripheral-display system to the operator's gaze position. A neural network controller is proposed that directly maps the gaze position to the narrow field projection co-ordinates without the need for any axis or co-ordinate transformations. A novel image feature-extraction algorithm, for extraction of the pupil-purkinje difference measure, has been developed that exhibits robust and reproducible real-time performance. By providing foveal and peripheral vision in a far-field teleoperator through the eye-directed split-foveal-peripheral-display, visual information is sufficiently and naturally provided for the establishment of telepresence. / Dissertation (M Eng (Electronic Engineering))--University of Pretoria, 2007. / Electrical, Electronic and Computer Engineering / unrestricted

This work presents the design, evaluation and comparison of evolution and reinforcement learning models, in isolation and combined in Darwinian and Lamarckian frameworks, with a particular emphasis being placed on their adaptive nature in response to environments that become increasingly unstable. Our ultimate objective is to determine whether hybrid models of evolution and learning can demonstrate adaptive qualities beyond those of such models when applied in isolation. This work demonstrates the limitations of evolution, reinforcement learning and Lamarckian models in dealing with increasingly unstable environments, while noting the effective adaptive nature of a Darwinian model to assimilate increasing levels of instability. This is shown to be a result of the Darwinian evolution model's ability to separate learning at two levels, the population's experience of the environment over the course of many generations and the individual's experience of the environment over the course of its lifetime. Thus, knowledge relating to the general characteristics of the environment over many generations can be maintained in the population's genotypes with phenotype (reinforcement) learning being utilized to adapt a particular agent to the particular characteristics of its environment. Lamarckian evolution, though, is shown to demonstrate adaptive characteristics that are highly effective in response to the stable environments. Selection and reproduction combined with reinforcement learning creates a model that has the ability to utilize useful knowledge produced by reinforcements, as opposed to random mutations, to accelerate the search process. As a result the influence of individual learning on the populations evolution is shown to be more successful when applied in the more direct Lamarckian form. Based on our results demonstrating the success of Lamarckian strategies in stable environments and Darwinian strategies in unstable environments, hybrid Darwinian/Lamarckian models are created with a view towards combining the advantages of both forms of evolution to produce a superior adaptive capability. Our investigation demonstrates that such hybrid models can effectively combine the adaptive advantageous of both Darwinian and Lamarckian evolution to provide a more effective capability of adapting to a range of conditions, from stable to unstable, appropriately adjusting the required degree of inheritance in response to the requirements of the environment.

Please read the abstract in the section 00front of this document / Dissertation (MSc)--University of Pretoria, 2005. / Computer Science / unrestricted

Application of EMG-controlled functional neuromuscular stimulation to a denervated muscle depends largely on the successful discrimination of the EMG signal by which the subject desires to execute control over the impeded movement. This can be achieved by an adaptive and flexible interface regardless of electrodes location, strength of remaining muscle activity or even personal conditions. Adaptability is a natural and important characteristic of artificial neural networks. This research work is restricted to the development of a real-time application of artificial neural network to the EMG signature recognition. Through this new approach, EMG features extracted by Fourier analysis are presented to a multilayer perceptron type neural network. The neural network learns the most relevant features of the control signal. For real-time operation, a digital signal processor operates over the resulting set of weights from the learning process, and maps the incoming signal to the stimulus control domain. Results showed a highly accurate discrimination of the EMG signal over interference patterns.

The importance of properly securing an organization’s information and computing resources has become paramount in modern business. Since the advent of the Internet, securing this organizational information has become increasingly difficult. Organizations deploy many security mechanisms in the protection of their data, intrusion detection systems in particular have an increasingly valuable role to play, and as networks grow, administrators need better ways to monitor their systems. Currently, many intrusion detection systems lack the means to accurately monitor and report on wireless segments within the corporate network. This dissertation proposes an extension to the NeGPAIM model, known as NeGPAIM-W, which allows for the accurate detection of attacks originating on wireless network segments. The NeGPAIM-W model is able to detect both wired and wireless based attacks, and with the extensions to the original model mentioned previously, also provide for correlation of intrusion attacks sourced on both wired and wireless network segments. This provides for a holistic detection strategy for an organization. This has been accomplished with the use of Fuzzy logic and neural networks utilized in the detection of attacks. The model works on the assumption that each user has, and leaves, a unique footprint on a computer system. Thus, all intrusive behaviour on the system and networks which support it, can be traced back to the user account which was used to perform the intrusive behavior.

Artificial neural networks are systems composed of interconnected simple computing units known as artificial neurons which simulate some properties of their biological counterparts. They have been developed and studied for understanding how brains function, and for computational purposes. In order to use a neural network for computation, the network has to be designed in such a way that it performs a useful function. Currently, the most popular method of designing a network to perform a function is to adjust the parameters of a specified network until the network approximates the input-output behaviour of the function. Although some analytical knowledge about the function is sometimes available or obtainable, it is usually not used. Some neural network paradigms exist where such knowledge is utilized; however, there is no systematical method to do so. The objective of this research is to develop such a method. A systematic method of neural network design, which we call algebraic derivation methodology, is proposed and developed in this thesis. It is developed with an emphasis on designing neural networks to implement image processing algorithms. A key feature of this methodology is that neurons and neural networks are represented symbolically such that a network can be algebraically derived from a given function and the resulting network can be simplified. By simplification we mean finding an equivalent network (i.e., performing the same function) with fewer layers and fewer neurons. A type of neural networks, which we call LQT networks, are chosen for implementing image processing algorithms. Theorems for simplifying such networks are developed. Procedures for deriving such networks to realize both single-input and multiple-input functions are given. To show the merits of the algebraic derivation methodology, LQT networks for implementing some well-known algorithms in image processing and some other areas are developed by using the above mentioned theorems and procedures. Most of these networks are the first known such neural network models; in the case there are other known network models, our networks have the same or better performance in terms of computation time. / Applied Science, Faculty of / Electrical and Computer Engineering, Department of / Graduate

Reinforcement learning algorithms comprise a class of learning algorithms for neural networks. Reinforcement learning is distinguished from other classes by the type of problems that it is intended to solve. It is used for learning input-output mappings where the desired outputs are not known and only a scalar reinforcement value is available. Primary Reinforcement Learning (PRL) is a core component of the most actively researched form of reinforcement learning. The issues surrounding the convergence characteristics of PRL are considered in this thesis. There have been no convergence proofs for any kind of networks learning under PRL. A convergence theorem is proved in this thesis, showing that under some conditions, a particular reinforcement learning algorithm, the A[formula omitted] algorithm, will train a single-layer network correctly. The theorem is demonstrated with a series of simulations. A new PRL algorithm is proposed to deal with the training of multiple layer, binary output networks with continuous inputs. This is a more difficult learning problem than with binary inputs. The new algorithm is shown to be able to successfully train a network with multiple outputs when the environment conforms to the conditions of the convergence theorem for a single-layer network. / Applied Science, Faculty of / Electrical and Computer Engineering, Department of / Graduate

In general, it is difficult to automatically detect faults within large scale engineering plants early during their onset. This is due to a number of factors including the large number of components typically present in such plants and the complex interactions of these components, which are typically poorly understood. Traditionally, fault detection within these plants has been performed through the use of status monitoring systems employing limit checking fault detection. In this approach, upper and lower bounds are placed on what is prescribed as “normal” behaviour for each of the plant's collected status data signals and fault flags are generated if and when the given status data signal exceeds either of its bounds. This approach tends to generate relatively large numbers of false alarms, due to the technique's inability to model known signal dependencies, and it also tends to produce inconsistent fault flags, in the sense that the flags do not tend to be produced throughout the “fault” event. The limit checking approach also is not particularly adept at early fault detection tasks since as long as the given status data signal remains between the upper and lower bounds any signal behaviour is deemed as acceptable. Hence, behavioural changes in the status data signals go undetected until their severity is such that either the upper or lower bounds are exceeded. In this dissertation, two novel fault detection methodologies are proposed which are better suited to the early fault detection task than traditional limit checking. The first technique is directed at modeling of signals exhibiting unknown linear dependencies. This detection system utilizes fuzzy membership functions to model signal behaviour and through this modelling approach fault detection bounds are generated which meet a prescribed probability of false alarm rate. The second technique is directed at modelling signals exhibiting unknown non-linear, dynamic dependencies. This system utilizes recurrent neural network technology to model the signal behaviours and prescribed statistical methods are employed to determine appropriate fault detection thresholds. Both of these detection systems have been designed to be able to be retrofitted into existing industrial status monitoring system and, as such, they have been designed to achieve good modelling performance in spite of the coarsely quantized status data signals which are typical of industrial status monitoring systems constructed to employ limit checking. The fault detection properties of the proposed fault detection systems were also compared to an in situ limit checking fault detection system for a set of real-world data obtained from an operational large scale engineering plant. This comparison showed that both of the proposed fault detection systems achieved marked improvements over traditional limit checking both in terms of their false alarm rates and their fault detection sensitivities. / Graduate

Magister Scientiae - MSc / Huge amounts of data are being collected as a result of the increased use of mobile telecommunications. Insight into information and knowledge derived from these databases can give operators a competitive edge in terms of customer care and retention, marketing and fraud detection. One of the strategies for fraud detection checks for signs of questionable changes in user behavior. Although the intentions of the mobile phone users cannot be observed, their intentions are reflected in the call data which define usage patterns. Over a period of time, an individual phone generates a large pattern of use. While call data are recorded for subscribers for billing purposes, we are making no prior assumptions about the data indicative of fraudulent call patterns, i.e. the calls made for billing purpose are unlabeled. Further analysis is thus, required to be able to isolate fraudulent usage. An unsupervised learning algorithm can analyse and cluster call patterns for each subscriber in order to facilitate the fraud detection process. This research investigates the unsupervised learning potentials of two neural networks for the profiling of calls made by users over a period of time in a mobile telecommunication network. Our study provides a comparative analysis and application of Self-Organizing Maps (SOM) and Long Short-Term Memory (LSTM) recurrent neural networks algorithms to user call data records in order to conduct a descriptive data mining on users call patterns. Our investigation shows the learning ability of both techniques to discriminate user call patterns; the LSTM recurrent neural network algorithm providing a better discrimination than the SOM algorithm in terms of long time series modelling. LSTM discriminates different types of temporal sequences and groups them according to a variety of features. The ordered features can later be interpreted and labeled according to specific requirements of the mobile service provider. Thus, suspicious call behaviours are isolated within the mobile telecommunication network and can be used to to identify fraudulent call patterns. We give results using masked call data from a real mobile telecommunication network. / South Africa

Artificial intelligence powered by deep neural networks has seen tremendous improvements in the last decade, achieving superhuman performance on a diverse range of tasks. Many worry that it can one day develop the ability to recursively self-improve itself, leading to an intelligence explosion known as the Singularity. Autogenerative networks, or neural networks generating neural networks, is one major plausible pathway towards realizing this possibility. The object of this thesis is to study various challenges and applications of small-scale autogenerative networks in domains such as artificial life, reinforcement learning, neural network initialization and optimization, gradient-based meta-learning, and logical networks. Chapters 2 and 3 describe novel mechanisms for generating neural network weights and embeddings. Chapters 4 and 5 identify problems and propose solutions to fix optimization difficulties in differentiable mechanisms of neural network generation known as Hypernetworks. Chapters 6 and 7 study implicit models of network generation like backpropagating through gradient descent itself and integrating discrete solvers into continuous functions. Together, the chapters in this thesiscontribute novel proposals for non-differentiable neural network generation mechanisms, significant improvements to existing differentiable network generation mechanisms, and an assimilation of different learning paradigms in autogenerative networks.