This thesis investigated the abilities of artificial neural networks as rock classifiers in an open-pit hard rock environment using monitoring-while-drilling (MWD) data. Blast hole drilling data has been collected from an open-pit taconite mine. The data was smoothed with respect to depth and filtered for non-drilling data. Preliminary analysis was performed to determine classifier input variables and a method of labelling training data. Results obtained from principal component analysis suggested that the best set of possible classifier input variables was: penetration rate, torque, specific fracture energy, vertical vibration, horizontal vibration, penetration rate deviation and thrust deviation. Specific fracture energy and self-organizing-maps were explored as a means of labelling training data and found to be inadequate. Several backpropagation neural networks were trained and tested with various combinations of input parameters and training sets. Input sets that included all seven parameters achieved the best overall performances. 7-input neural networks that were trained with and tested on the entire data set achieved an average overall performance of 81%. A sensitivity analysis was performed to test the generalization abilities of the neural networks as rock classifiers. The best overall neural network performance on data not included in the training set was 67%. The results indicated that neural networks by themselves are not capable rock classifiers on MWD data in such a hard rock iron ore environment. / Thesis (Master, Mining Engineering) -- Queen's University, 2009-04-23 11:59:07.806

We explore the training of deep neural networks to produce vector representations using weakly labelled information in the form of binary similarity labels for pairs of training images. Previous methods such as siamese networks, IMAX and others, have used fixed cost functions such as $L_1$, $L_2$-norms and mutual information to drive the representations of similar images together and different images apart. In this work, we formulate learning as maximizing the likelihood of binary similarity labels for pairs of input images, under a parameterized probabilistic similarity model. We describe and evaluate several forms of the similarity model that account for false positives and false negatives differently. We extract representations of MNIST, AT\&T ORL and COIL-100 images and use them to obtain classification results. We compare these results with state-of-the-art techniques such as deep neural networks and convolutional neural networks. We also study our method from a dimensionality reduction prospective.

We explore the training of deep neural networks to produce vector representations using weakly labelled information in the form of binary similarity labels for pairs of training images. Previous methods such as siamese networks, IMAX and others, have used fixed cost functions such as $L_1$, $L_2$-norms and mutual information to drive the representations of similar images together and different images apart. In this work, we formulate learning as maximizing the likelihood of binary similarity labels for pairs of input images, under a parameterized probabilistic similarity model. We describe and evaluate several forms of the similarity model that account for false positives and false negatives differently. We extract representations of MNIST, AT\&T ORL and COIL-100 images and use them to obtain classification results. We compare these results with state-of-the-art techniques such as deep neural networks and convolutional neural networks. We also study our method from a dimensionality reduction prospective.

This thesis is concerned with the use of neural networks for motor control tasks. The main goal of the thesis is to investigate ways in which the biological notions of motor programs and Central Pattern Generators (CPGs) may be implemented in a neural network framework. Biological CPGs can be seen as components within a larger control scheme, which is basically modular in design. In this thesis, these ideas are investigated through the use of modular recurrent networks, which are used in a variety of control tasks. The first experimental chapter deals with learning in recurrent networks, and it is shown that CPGs may be easily implemented using the machinery of backpropagation. The use of these CPGs can aid the learning of pattern generation tasks; they can also mean that the other components in the system can be reduced in complexity, say, to a purely feedforward network. It is also shown that incremental learning, or 'shaping' is an effective method for building CPGs. Genetic algorithms are also used to build CPGs; although computational effort prevents this from being a practical method, it does show that GAs are capable of optimising systems that operate in the context of a larger scheme. One interesting result from the GA is that optimal CPGs tend to have unstable dynamics, which may have implications for building modular neural controllers. The next chapter applies these ideas to some simple control tasks involving a highly redundant simulated robot arm. It was shown that it is relatively straightforward to build CPGs that represent elements of pattern generation, constraint satisfaction. and local feedback. This is indirect control, in which errors are backpropagated through a plant model, as well as the ePG itself, to give errors for the controller. Finally, the third experimental chapter takes an alternative approach, and uses direct control methods, such as reinforcement learning. In reinforcement learning, controller outputs have unmodelled effects; this allows us to build complex control systems, where outputs modulate the couplings between sets of dynamic systems. This was shown for a simple case, involving a system of coupled oscillators. A second set of experiments investigates the use of simplified models of behaviour; this is a reduced form of supervised learning, and the use of such models in control is discussed.

Much experimental evidence indicates that lat perceptual decisions are made by integrating sensory information in cortical areas, until the accumulated evidence fies certain criteria. Recently proposed theories further suggest that the ain performs statistically optimal strategies during decision processes. This thesis extends and develops biologically inspired decision models from different aspects.

No description available.

This thesis deals with automatic defect detection. The objective was to develop the techniques required by a small manufacturing business to make cost-efficient use of inspection technology. In our work on inspection techniques we discuss image acquisition and the choice between custom and general-purpose processing hardware. We examine the classes of general-purpose computer available and study popular operating systems in detail. We highlight the advantages of a hybrid system interconnected via a local area network and develop a sophisticated suite of image-processing software based on it. We quantitatively study the performance of elements of the TCP/IP networking protocol suite and comment on appropriate protocol selection for parallel distributed applications. We implement our own distributed application based on these findings. In our work on inspection algorithms we investigate the potential uses of iterated function series and Fourier transform operators when preprocessing images of defects in aluminium plate acquired using a linescan camera. We employ a multi-layer perceptron neural network trained by backpropagation as a classifier. We examine the effect on the training process of the number of nodes in the hidden layer and the ability of the network to identify faults in images of aluminium plate. We investigate techniques for introducing positional independence into the network's behaviour. We analyse the pattern of weights induced in the network after training in order to gain insight into the logic of its internal representation. We conclude that the backpropagation training process is sufficiently computationally intensive so as to present a real barrier to further development in practical neural network techniques and seek ways to achieve a speed-up. Weconsider the training process as a search problem and arrive at a process involving multiple, parallel search "vectors" and aspects of genetic algorithms. We implement the system as the mentioned distributed application and comment on its performance.

No description available.

This thesis investigates the way in which information about the structure of a set of training data with 'natural' characteristics may be used to positively influence the design of associative memory neural network models of the Hopfield type. This is done with a view to reducing the level of connectivity in models of this type. There are three strands to this work. Firstly, an empirical evaluation of the implementation of existing theory is given. Secondly, a number of existing theories are combined to produce novel network models and training regimes. Thirdly, new strategies for constructing and training associative memories based on knowledge of the structure of the training data are proposed. The first conclusion of this work is that, under certain circumstances, performance benefits may be gained by establishing the connectivity in a non-random fashion, guided by the knowledge gained from the structure of the training data. These performance improvements exist in relation to networks in which sparse connectivity is established in a purely random manner. This dilution occurs prior to the training of the network. Secondly, it is verified that, as predicted by existing theory, targeted post-training dilution of network connectivity provides greater performance when compared with networks in which connections are removed at random. Finally, an existing tool for the analysis of the attractor performance of neural networks of this type has been modified and improved. Furthermore, a novel, comprehensive performance analysis tool is proposed.

No description available.