Advances in neuroscience implicate reentrant signaling as the predominant form of communication between brain areas. This principle was used in a series of masking experiments that defy explanation by feed-forward theories. The masking occurs when a brief display of target plus mask is continued with the mask alone. Two masking processes were found: an early process affected by physical factors such as adapting luminance and a later process affected by attentional factors such as set size. This later process is called masking by object substitution, because it occurs whenever there is a mismatch between the reentrant visual representation and the ongoing lower level activity. Iterative reentrant processing was formalized in a computational model that provides an excellent fit to the data. The model provides a more comprehensive account of all forms of visual masking than do the long-held feed-forward views based on inhibitory contour interactions. From the time a stimulus first enters the eye to the time a percept emerges into consciousness, the initial stimulus has been coded at several levels in the visual system. One of the main goals in studying visual information processing is to specify the representations at each level and the temporal sequence between

Abstract- This paper discusses Memory Neuron Networks as models for identification and adaptive control of nonlinear dy-namical systems. These are a class of recurrent networks obtained by adding trainable temporal elements to feed-forward networks that makes the output history-sensitive. By virtue of this capa-bility, these networks can identify dynamical systems without having to be explicitly fed with past inputs and outputs. Thus, they can identify systems whose order is unknown or systems with unknown delay. It is argued that for satisfactory modeling of dynamical systems, neural networks should be endowed with such internal memory. The paper presents a preliminary analysis of the learning algorithm, providing theoretical justification for the identification method. Methods for adaptive control of nonlinear systems using these networks are presented. Through extensive simulations, these models are shown to be effective both for identification and model reference adaptive control of nonlinear systems. I.

In order to take advantage of the massive parallelism offered by artificial neural networks, hardware implementations are essential. However, most standard neural network models are not very suitable for implementation in hardware and adaptations are needed. In this section an overview is given of the various issues that are encountered when mapping an ideal neural network model onto a compact and reliable neural network hardware implementation, like quantization, handling non-uniformities and non-ideal responses, and restraining computational complexity. Furthermore, a broad range of hardware-friendly learning rules is presented, which allow for simpler and more reliable hardware implementations. The relevance of these neural network adaptations to hardware is illustrated by their application in existing hardware implementations.

Neural networks are usually trained using local, gradient-based procedures. Such methods frequently find suboptimal solutions being trapped in local minima. Optimization of neural structures and global minimization methods applied to network cost functions have strong influence on all aspects of network performance. Recently genetic algorithms are frequently combined with neural methods to select best architectures and avoid drawbacks of local minimization methods. Many other global minimization methods are suitable for that purpose, although they are used rather rarely in this context. This paper provides a survey of such global methods, including some aspects of genetic algorithms.

This paper reviews the techniques that reduce the time complexity and improve the convergence capability of the real-time recurrent learning algorithm. A comparison among the various approaches was made by training several recurrent networks to model a chaotic time series produced by the Henon model. 1. INTRODUCTION The real-time recurrent learning (RTRL) algorithm [1] is one of the successful learning algorithms where the gradient of errors is propagated forward in time. Therefore, it is particularly suitable for on-line training of recurrent neural networks (RNNs). Nevertheless, its time complexity is O(n 4 ), where n is the number of processing units in the network. After its introduction in 1989, a number of suggestions have been made to improve the learning speed and convergence of the algorithm. 2. METHODS TO IMPROVE THE RTRL ALGORITHM Before the improved RTRL algorithms are discussed, we need to define the original RTRL algorithm [1]. Let the parameters of a fully connecte...

Algorithms based on systematic search techniques can be successfully applied for multilayer perceptron (MLP) training and for logical rule extraction from data using MLP networks. The proposed solutions are easier to implement and frequently outperform gradient-based optimization algorithms. Search-based techniques, popular in artificial intelligence and almost completely neglected in neural networks can be the basis for MLP network training algorithms. There are plenty of well-known search algorithms, however since they are not suitable for MLP training, new algorithms dedicated to this task must be developed. Search algorithms applied to MLP networks change network parameters (weights and biases) and check the influence of the changes on the error function. MLP networks considered in this thesis are used for data classification and logical rule-based understanding of the data. The proposed solutions in many cases outperform gradient-based backpropagation algorithms. The thesis is organized in three parts. The first part of the thesis concentrates on better understanding of MLP properties.

Neural networks are usually trained using local, gradient-based procedures, and the best architectures are selected by experimentation. Gradient methods frequently find suboptimal solutions being trapped in local minima. Genetic algorithms are frequently used but do not guarantee optimal solutions and are computationally expensive. Several new global optimization methods suitable for architecture optimization and neural training are described here. Multistart initialization methods are also offered as an alternative to global minimization. I. INTRODUCTION S OFT computing methods compete with traditional pattern recognition and statistical methods in many applications. For neural networks with predetermined structure, for example Multilayer Perceptrons (MLPs) with fixed architectures, finding an optimal set of parameters (weights and thresholds) requires a solution of a non-linear optimization problem. Such problems in general are NP-complete and the chance to find the best solu...

A feedforward network is proposed which lends itself to cost-e#ective implementations in digital hardware and has a fast forward-pass capability. It di#ers from the conventional model in restricting its synapses to the set 0, 1} while allowing unrestricted o#sets. Simulation results on the `onset of diabetes' data set and a handwritten numeral recognition database indicate that the new network, despite having strong constraints on its synapses, has a generalization performance similar to that of its conventional counterpart.

Weight-perturbation (WP) algorithms for supervised and/or reinforcement learning offer improved biological plausibility over backpropagation because of their reduced circuitry requirements for realization in neural hardware. All such algorithms use some form of information source — a means to compare weight changes with changes in output error — to adjust weights. This paper explores the hypothesis that biological synaptic noise might serve as the substrate by which weight perturbation is implemented. We explore the basic synaptic noise hypothesis (BSNH) which embodies the weakest assumptions about the underlying neural circuitry required to implement WP algorithms. The present paper identifies relevant biological constraints consistent with the BSNH, taxonomizes existing WP algorithms in regard to consistency with those constraints, and proposes a new WP algorithm that is fully consistent with the constraints. By comparing the learning effectiveness of these algorithms via simulation studies, we find that all of the algorithms can support traditional neural network learning tasks and have similar generalization characteristics, although the results suggest a trade-off between learning efficiency and biological accuracy. This establishes the basic result that biological synaptic noise, coupled with appropriate reward, can be exploited to implement WP algorithms for neural network learning. 1

This thesis is dedicated to my family, and especially my mother. Acknowledgements In writing this thesis, as the completion of a year of hard work, I have many people to thank for their help along the way. I wish to thank my supervisor, Professor John W. Lloyd, for helping to guide my project throughout the many twists and turns it has taken from random ideas into a concrete research project, and for helping me transform those ideas into reality. I also thank him for his patience in helping proof read the many drafts of this thesis. I wish to thank Kee Siong Ng for having laid many steps along the way which have allowed me to focus on other aspects of this project, and for some insightful comments while proof reading this thesis. I wish to thank my mother and older sister for their constant encouragement and support throughout this past year. I would also like to thank John Uhlmann for many useful discussions and very careful proof-reading of my thesis, David Hellier for picking at my grammar, Richard Walker