J. M. Bishop

The Stochastic Search Network

A fundamental difficulty when using neural networks applied to pattern recognition, is the problem of stimulus equivalence — the invariance of symbolic information independent of transformation within a search space. For example, symbolically, the letter A remains an A irrespective of its position, size or orientation within an image field. Adult humans can generally recognise patterns accurately, despite such transformations and distortions. Classically it has been hypothesised that this ability is due to a normalisation process that occurs before the classification process begins.

Self-Localisation in the ‘Senario’ Autonomous Wheelchair

This paper introduces the Focused Stochastic Diffusion Network as a novel method of self-localisation for an autonomous wheelchair in a complex, busy environment. The space of possible positions is explored in parallel by a set of cells searching in a competitive co-operative manner for the most likely position of the wheelchair in its environment. Experimental results from the SENARIO autonomous wheelchair project indicate the technique is practical and robust.

Kubelka-Munk or neural networks for computer colorant formulation?

Traditionally Computer Colorant Formulation has been implemented using a theory of radiation transfer known as Kubelka-Munk (K-M) theory. Kubelka-Munk theory allows the prediction of spectral reflectance for a mixture of components (colorants) that have been characterised by absorption K and scattering S coefficients. More recently it has been suggested that Artifical Neural Networks ANNs) may be able to provide alternative mappings between colorant concentrations and spectral reflectances and, more generally, are able to provide transforms between color spaces. This study investigates the ability of ANNs to predict spectral reflectance from colorant concentrations using a set of data measured from known mixtures of lithographic printing inks. The issue of over-training is addressed and we show that the number of hidden units in the network must be carefully selected. We show that it is difficult to train a conventional neural network to the level that matches the performance that can be achieved using the K-M theory. However, a hybrid model is proposed that may out-perform the K-M model.

Virtual bodies and virtual spaces

In the Transcendental Aesthetic part of the Critique of Pure Reason, Immanuel Kant stated the a priori necessity of the singularity of space that, “we can represent to ourselves only one space; and if we speak of diverse spaces, we mean thereby only parts of one and the same space … these parts cannot precede the one all‐embracing space … they can be thought only as in it”. If correct, Kant places a tight bound around the universe we consciously inhabit. Established arguments against Kant’s claims are reviewed and criticised based on the notion of dream spaces, before outlining the novel hypothesis that the widespread use of cyberspace and large scale multi‐user virtual realities illustrate public spaces beyond physical reality, and as such provide an empirical refutation of the a priori necessity of the singularity of space.

Computer‐mediated communication use by the deaf and hard‐of‐hearing

Discusses the potential of computer‐mediated communication to reduce the social isolation experienced by many deaf and hard‐of‐hearing individuals. Communication presents significant problems for this group of people, some of which can be bridged by communicating via the Internet or e‐mail. deaf Internet users were surveyed by use of a questionnaire. Their opinions, summarised in this paper, emphasise the significance of computer‐mediated communication for the deaf; namely that interaction can be both less stressful and faster, thereby reducing the isolation many feel due to the physical constraints imposed by their deafness.

Genetic Optimisation of Neural Network Architectures for Colour Recipe Prediction

Colour control systems based on spectrophotometers and microprocessors are finding increased use in production environments. One of the most important aspects of quality control in manufacturing processes is the maintenance of colour in the product. This involves selecting a recipe of appropriate dyes or pigments which when applied at a specific concentration to the product will render the required colour. This process is known as recipe prediction and is traditionally carried out by trained colourists who achieve a colour match via a combination of experience and trial-and-error. Instrumental recipe prediction was introduced commercially in the 1960’s and has become one of the most important industrial applications of colorimetry. The model that is almost exclusively used is known as the Kubelka-Munk theory, however its operation in certain areas of coloration is such as to warrant an alternative approach. The purpose of this paper is to investigate the performance of a Genetically optimised Neural Network applied to this recipe prediction task.

Stochastic diffusion search

Stochastic Diffusion Search is a well characterised robust swarm intelligence global metaheuristic, that can efficiently solve search and optimisation problems with compositional structure

Neural Networks in the Colour Industry

In the past ten years there has been an explosion of academic interest in Neural Network research, yet the techniques are still viewed with some suspicion by many engineers faced with real world problems. The purpose of this paper is to illustrate how a simple neural network is being used to help solve a difficult physical problem. The work, sponsored by Courtaulds Research, involves colour recipe prediction. It is a difficult problem to solve using conventional computer techniques as the model that is most widely used (Kubelka-Munk theory) breaks down under a variety of conditions. The paper will discuss several of the design decisions, common to many neural network applications, that have been made in the process of developing the Courtaulds Recipe Prediction System.

Prediction of Reflectance Values: Towards the Integration of Neural and Conventional Colorimetry

Colour control systems based on spectrophotometers and microprocessors are finding increased use in production environments. One of the most important aspects of quality control in manufacturing processes is the maintenance of colour in the product This involves selecting a recipe of appropriate dyes or pigments which when applied at a specific concentration to the product will render the required colour. This process is known as recipe prediction and is traditionally carried out by trained colourists who achieve a colour match via a combination of experience and trial-and-error. Instrumental recipe prediction was introduced commercially in the 1960 ‘s and has become one of the most important industrial applications of colorimetry. The model that is almost exclusively used is known as the Kubelka-Munk theory, however its operation in certain areas of coloration is such as to warrant an alternative approach. Previous papers [10, 11] have demonstrated the feasibility of a neural network method of predicting dye recipes from a colour specification, however these techniques have not been directly applicable to commercial systems. This paper will describe new research at Keele that enables neural techniques to be simply embodied into current recipe prediction systems.

A Comparison Between Chemotaxis and Back-Propagation Learning Applied to Colour Recipe Prediction.

Abstract Conventional computer colour recipe prediction systems employ optical models (commonly the Kubelka-Munk theory) to relate measured reflectance values to colorant concentration. However these systems provide only an approximate model and hence situations exist where this approach is not applicable. An alternative method is to use Artificial Intelligence techniques. Initial research into the use of neural networks for colour recipe prediction used a multi-layer perceptron architecture and back-propagation learning. Although back-propagation works well, several criticisms have emerged concerning its speed, biological feasibility and ease of VLSI implementation. The chemotaxis algorithm promises to be efficient and easily implementable in VLSI. The paper will describe the chemotaxis algorithm, the ways in which it overcomes some of the problems inherent in back-propagation and compare its performance with the back-propagation algorithm with reference to the colour recipe problem.