Pattern recognition is a rapidly expanding area of research, with applications ranging from character recognition and component inspection to robotic guidance and military reconnaissance. The basic principle of image recognition is that of comparing the unknown image with many known reference images or 'filters', until a match is found. By comparing the unknown image with a large data bank of filters, the diversity of the application can be extended. The work presented in this thesis details the practical development of an optical disk based memory system as applied in various optical correlators for pattern recognition purposes. The characteristics of the holographic optical disk as a storage medium are investigated in terms of information capacity and signal to noise ratio, where a fully automated opto-mechanical system has been developed for the control of the optical disk and the processing of the information recorded. A liquid crystal television has been used as a Spatial Light Modulator for inputting the image data, and as such, the device characteristics have been considered with regard to processing both amplitude and phase information. Three main configurations of optical correlator have been applied, specifically an image plane correlator, a VanderLugt correlator, and an Anamorphic correlator. Character recognition has been used to demonstrate correlator performance, where simple matched filtering has been applied, subsequent to which, an improvement in class discrimination has been demonstrated with the application of the Minimum Average Correlation Energy filter. The information processing rate obtained as a result of applying 2D parallel processing has been shown to be many orders of magnitude larger than that available with comparable serial based digital systems.

Thesis (Ph.D.)--University of Melbourne, Howard Florey Institute, 2005. / Typescript. Includes bibliographical references (leaves 200-272).

The values of a given manipulator's dynamics coefficients need to be accurately identified in order to employ model-based algorithms in the control of its motion. This thesis details the development of a novel form of adaptive network which is capable of accurately learning the coefficients of systems, such as manipulator inverse dynamics, where the algebraic form is known but the coefficients' values are not. Empirical motion data from a pair of PUMA 560s has been processed by the Context-Sensitive Linear Combiner (CSLC) network developed, and the coefficients of their inverse dynamics identified. The resultant precision of control is shown to be superior to that achieved from employing dynamics coefficients derived from direct measurement. As part of the development of the CSLC network, the process of network learning is examined. This analysis reveals that current network architectures for processing analogue output systems with high input order are highly unlikely to produce solutions that are good estimates throughout the entire problem space. In contrast, the CSLC network is shown to generalise intrinsically as a result of its structure, whilst its training is greatly simplified by the presence of only one minima in the network's error hypersurface. Furthermore, a fine-tuning algorithm for network training is presented which takes advantage of the CSLC network's single adaptive layer structure and does not rely upon gradient descent of the network error hypersurface, which commonly slows the later stages of network training.

This thesis aims to understand how to design high performance, flexible and cost effective neural computing systems and apply them to a variety of real-time applications. Systems of this type already exist for the support of a range of ANN models. However, many of these designs have concentrated on optimising the architecture of the neural processor and have generally neglected other important aspects. If these neural systems are to be of practical benefit to researchers and allow complex neural problems to be solved efficiently, all aspects of their design must be addressed.

Modular connectionist systems comprise autonomous, communicating modules, achieving a behaviour more complex than that of a single neural network. The component modules, possibly of different topologies, may operate under various learning algorithms. Some modular connectionist systems are constrained at the representational level, in that the connectivity of the modules is hard-wired by the modeller; others are constrained at an architectural level, in that the modeller explicitly allocates each module to a specific subtask. Our approach aims to minimise these constraints, thus reducing the bias possibly introduced by the modeller. This is achieved, in the first case, through the introduction of adaptive connection weights and, in the second, by the automatic allocation of modules to subtasks as part of the learning process. The efficacy of a minimally constrained system, with respect to representation and architecture, is demonstrated by a simulation of numerical development amongst children. The modular connectionist system MASCOT (Modular Architecture for Subitising and Counting Over Time) is a dual-routed model simulating the quantification abilities of subitising and counting. A gating network learns to integrate the outputs of the two routes in determining the final output of the system. MASCOT simulates subitising through a numerosity detection system comprising modules with adaptive weights that self-organise over time. The effectiveness of MASCOT is demonstrated in that the distance effect and Fechner's law for numbers are seen to be consequences of this learning process. The automatic allocation of modules to subtasks is illustrated in a simulation of learning to count. Introducing feedback into one of two competing expert networks enables a mixture-of-experts model to perform decomposition of a task into static and temporal subtasks, and to allocate appropriate expert networks to those subtasks. MASCOT successfully performs decomposition of the counting task with a two-gated mixture-of-experts model and exhibits childlike counting errors.

This thesis is about adaptive invariance, and a new model of it: the Group Representation Network. We begin by discussing the concept of adaptive invariance. We then present standard background material, mostly from the fields of group theory and neural networks. Following this we introduce the problem of invariant pattern recognition and describe a number of methods for solving various instances of it. Next, we define the Symmetry Network, a connectionist model of permutation invariance, and we develop some new theory of this model. We also extend the applicability of the Symmetry Network to arbitrary finite group actions. We then introduce the Group Representation Network (GRN) as an abstract model, with which in principle we can construct concomitants between arbitrary group representations. We show that the GRN can be regarded as a neural network model, and that it includes the Symmetry Network as a submodel. We apply group representation theory to the analysis of GRNs. This yields general characterizations of the allowable activation functions in a GRN and of their weight matrix structure. We examine various generalizations and restricted cases of the GRN model, and in particular look at the construction of GRNs over infinite groups. We then consider the issue of a GRN's discriminability, which relates to the problem of graph isomorphism. We look next at the computational abilities of the GRN, and postulate that it is capable of approximately computing any group concomitant. We show constructively that any polynomial concomitant can be computed by a GRN. We also prove that a variety of standard models for invariant pattern recognition can be viewed as special instances of the GRN model. Finally, we propose that the GRN model may be biologically plausible and give suggestions for further research.

No description available.

Many consumer facing companies have large scale data sets that they use to create recommendations for their users. These recommendations are usually based off information the company has on the user and on the item in question. Based on these two sets of features, models are created and tuned to produce the best possible recommendations. A third set of data that exists in most cases is the presence of past interactions a user may have had with other items. The relationships that a model can identify between this information and the other two types of data, we believe, can improve the prediction of how a user may interact with the given item. We propose a method that can inform the model of these relationships during the training phase while only relying on the user and item data during the prediction phase. Using ideas from convolutional neural networks (CNN) and collaborative filtering approaches, our method manipulated the weights in the first layer of our network design in a way that achieves this goal.

No description available.

No description available.