Search results

Reviews some of the good reasons for looking to real neural nets for guidance on ways of implementing effective parallel computation. Discusses existing artificial neural nets with particular attention to the extent to which they model real neural activity. Indicates some serious mismatches, but shows that there are also important correspondences. The successful applications are to image processing, pattern classification and automatic optimization, in various guises. Reviews important issues raised by extension to the symbolic problem solving of “intellectual” thought, the prime concern of classical AI. These illustrate the importance of recursion, and of a degree of continuity associated with any evolutionary process.

The purpose of this paper is to outline an integrative, high-level, neurocomputational theory of brain function based on temporal codes, neural timing nets, and active regeneration of temporal patterns of spikes within recurrent neural circuits that provides a time-domain alternative to connectionist approaches.

This conceptual-theoretical paper draws from cybernetics, theoretical biology, neurophysiology, integrative and computational neuroscience, psychology, and consciousness studies.

The high-level functional organization of the brain involves adaptive cybernetic, goal-seeking, switching, and steering mechanisms embedded in percept-action-environment loops. The cerebral cortex is conceived as a network of reciprocally connected, re-entrant loops within which circulate neuronal signals that build up, decay, and/or actively regenerate. The basic signals themselves are temporal patterns of spikes (temporal codes), held in the spike correlation mass-statistics of both local and global neuronal ensembles. Complex temporal codes afford multidimensional vectorial representations, multiplexing of multiple signals in spike trains, broadcast strategies of neural coordination, and mutually reinforcing, autopoiesis-like dynamics. Our working hypothesis is that complex temporal codes form multidimensional vectorial representations that interact with each other such that a few basic processes and operations may account for the vast majority of both low- and high-level neural informational functions. These operational primitives include mutual amplification/inhibition of temporal pattern vectors, extraction of common signal dimensions, formation of neural assemblies that generate new temporal pattern primitive “tags” from meaningful, recurring combinations of features (perceptual symbols), active regeneration of temporal patterns, content-addressable temporal pattern memory, and long-term storage and retrieval of temporal patterns via a common synaptic and/or molecular mechanism. The result is a relatively simplified, signal-centric view of the brain that utilizes universal coding schemes and pattern-resonance processing operations. In neurophenomenal terms, waking consciousness requires regeneration and build up of temporal pattern signals in global loops, whose form determines the contents of conscious experience at any moment.

Understanding how brains work as informational engines has manifold long-reaching practical implications for design of autonomous, adaptive robotic systems. By proposing how new concepts might arise in brains, the theory bears potential implications for constructivist theories of mind, i.e. how observer-actors interacting with one another can self-organize and complexify.

The theory is highly original and heterodox in its neural coding and neurocomputational assumptions. By providing a possible alternative to standard connectionist theory of brain function, it expands the scope of thinking about how brains might work as informational systems.

The usefulness of artificial neural nets stems from their ability to self‐adjust, or in some sense “learn”. In modern studies, the emphasis on powerful self‐organisation is less strong, but the early viewpoint is defended here as potentially useful. Possible extension of neural net capability to “symbolic” processing is related to Minsky’s “heuristic connection” and to Pask’s view of learning as necessarily involving reformulation of information in a new language. Relevance is demonstrated to the “Boxes” scheme of Michie and Chambers and recent developments in reinforcement learning.

Quality assessment and fault detection are important topics in textile research. Human assessment in this field, however, is subjective and slow. Presents an automatic assessment using two fundamentally different kinds of neural networks: the Kohonen Map (an unsupervised system) and the backpropagation network (supervised system). Evaluates two case studies using these techniques: assessment of carpet wear and the assessment of set marks. Both show good results when applied to the aforementioned problems. Makes a comparison between the two techniques and shows that the unsupervised system also gives an evaluation of the objectivity of the human experts.

Knowledge of the microwave transistor parameters at various bias conditions is often required in computer‐aided design of complex microwave low‐noise circuit. Since the measurements of noise parameters are very complex and time‐consuming, microwave circuit designers usually use the catalogues' data or noise models. The noise data that can be found in the catalogues are often limited to a few frequencies and to one or few bias points. Further, most of the existing noise models require recalculation of elements/parameters of an equivalent circuit for every bias point. Microwave HEMT transistor noise prediction based on a multilayer perceptron neural network, proposed in this paper, enables noise prediction for all operating points over a wide frequency range. Neural networks are trained to learn noise parameters' dependence on bias conditions and frequency. After network training, noise prediction for a specified bias point requires only a network response calculation without changes in the network structure.

The application of an artificial neural network (ANN) to forecast the construction duration of buildings at the predesign stage is described in this paper. A three‐layered back‐propagation (BP) network consisting of 11 input nodes has been constructed. Ten binary input nodes represent basic information on building features (i.e. building function, structural system, foundation, height, exterior finishing, quality of interior decorating, and accessibility to the site), and one real‐value input represents functional area. The input nodes are fully connected to one output node through hidden nodes. The network was implemented on a Pentium‐150 based microcomputer using a neurocomputer program written in C+ +. The Generalized Delta Rule (GDR) was used as learning algorithm. One hundred and thirty‐six buildings built during the period 1987–95 in the Greater Bangkok area were used for training and testing the network. The determination of the optimum number of hidden nodes, learning rate, and momentum were based on trial‐and‐error. The best network was found to consist of six hidden nodes, with a learning rate of 0.6, and null momentum. It was trained for 44700 epochs within 943 s such that the mean squared error (judgement) of training and test samples were reduced to 1.17 × 10−7 and 3.10 × 10−6, respectively. The network can forecast construction du‐ration at the predesign stage with an average error of 13.6%.

The evolution of the concept of an Adaptive Network Library Interface is described and several technical and research issues are explored. The Adaptive Network Library Interface (ANLI) is a computer program that stands as a buffer between users of the library catalog and the catalog itself. This buffer unit maintains its own network of pointers from book to book, which it elicits from the users, interactively. It is hoped that such a buffer increases the value of the catalog for the users and provides librarians with new and useful information about the books in the collection. The relation to concepts such as hypertext and neural networks is explored as well.

The purpose of this paper is to develop a methodology for the design of cellular neural networks with interconnection topologies optimized and suitable for spatially distributed implementation.

The authors perform combinatorial optimization on the neural network's topology to obtain a sparser network, in which the links between the components of the network that reside in different physical locations are minimized. The approach builds on existing computationally efficient tools for the design of cellular neural networks and uses the concept of the network's stability parameters to assess the performance of the network prior to testing.

It turns out that the sparser cellular neural networks thus produced exhibit performance that can be on par with that of networks with full connectivity, and that for implementations of modest size, communication delays are not that significant to affect the stability of the dynamical system.

The novelty of the proposed approach lies in the formulation of the combinatorial optimization problem in a way that trades‐off network performance for communication overhead, and the use of this method for the physical implementation of associative memories across different interconnected processors.

An artificial neural network methodology is used to develop a new measure of contagion using exchange rate data from the Asian Crisis of 1997 and beyond. Connection weight changes during retraining of networks used to forecast exchange rates form the basis of this measure. These weight changes are used in obtaining a contribution factor for independent variables used in a forecasting process. Volatilities of contribution factors form the basis of the measure of contagion obtained. These volatilities are statistically validated through a series of simulations where critical values for them are derived. The measures of contagion obtained are then matched to concurrent economic and financial shocks that occurred during the crisis. It is found that there is good correlation between these events and the contagion measures obtained.