Search

A novel method is presented to allow a machine controller to evolve while the machine is acting in its environment. The method uses a single spiking neural network with a minimum number of neurons and no initial connections. New connections and neurons are grown by evaluating reward values which can represent either the internal state of the machine or the rating of its task performance. This way the topology and the level of connectivity of the network are kept to a minimum. The method will be applied to a controller for an autonomous mobile robot. [ABSTRACT FROM AUTHOR]

The Recursive Deterministic Perceptron (RDP) feed-forward multilayer neural network is a generalisation of the single layer perceptron topology. This model is capable of solving any two-class classification problem unlike the single layer perceptron which can only solve classification problems dealing with linearly separable sets. For all classification problems, the construction of an RDP is done automatically and convergence is always guaranteed. A generalisation of the 2-class Recursive Deterministic Perceptron (RDP) exists. This generalisation always allows the deterministic separation of m-classes. It is based on a new notion of linear separability and it arises naturally from the 2 valued RDP. The methods for building 2-class RDP neural networks have been extensively tested. However, no testing has been done before on the m-class RDP method. For the first time, a study on the performance of the m-class method is presented. This study will allow the highlighting of the main advantages and disadvantages of this method by comparing the results obtained while building m-class RDP neural networks with other more classical methods such as Backpropagation and Cascade Correlation in terms of level of generalisation and topology size. The networks were trained and tested using the following standard benchmark classification datasets: Glass, Wine, Zoo, Iris, Soybean, and Wisconsin Breast Cancer. [ABSTRACT FROM AUTHOR]

This chapter presents and discusses several well-known constructive neural network algorithms suitable for constructing feedforward architectures aiming at classification tasks involving two classes. The algorithms are divided into two different groups: the ones directed by the minimization of classification errors and those based on a sequential model. In spite of the focus being on two-class classification algorithms, the chapter also briefly comments on the multiclass versions of several two-class algorithms, highlights some of the most popular constructive algorithms for regression problems and refers to several other alternative algorithms. [ABSTRACT FROM AUTHOR]

This paper presents a novel neural network approach to recovering of 3D surfaces from single gray scale images. The proposed neural network uses photometric stereo to estimate local surfaces orientation for surfaces at each point of the surface that was observed from same viewpoint but with different illumination direction in surfaces that follow the Lambertian reflection model. The parameters for the neural network are a 3x3 brightness patch with pixel values of the image and the light source direction. The light source direction of the surface is calculated using two different approaches. The first approach uses a mathematical method and the second one a neural network method. The images used to test the neural network were both synthetic and real images. Only synthetic images were used to compare the approaches mainly because the surface was known and the error could be calculated. The results show that the proposed neural network is able to recover the surface with a highly accurately estimate. [ABSTRACT FROM AUTHOR]

The Recursive Deterministic Perceptron (RDP) feed-forward multilayer neural network is a generalisation of the single layer perceptron topology. This model is capable of solving any two-class classification problem as opposed to the single layer perceptron which can only solve classification problems dealing with linearly separable sets. For all classification problems, the construction of an RDP is done automatically and convergence is always guaranteed. A generalization of the 2-class Recursive Deterministic Perceptron (RDP) exists. This generalization allows to always separate, in a deterministic way, m-classes. It is based on a new notion of linear separability and it falls naturally from the 2 valued RDP. The methods for building 2-class RDP neural networks have been extensively tested. However, no testing has been done before on the m-class RDP method. For the first time, a study on the performance of the m-class method is presented. This study will allow the highlighting of the main advantages and disadvantages of this method by comparing the results obtained while building m-class RDP neural networks with other more classical methods such as Backpropagation and Cascade Correlation. The networks were trained and tested using the following standard benchmark classification datasets: IRIS, SOYBEAN, and Wisconsin Breast Cancer. [ABSTRACT FROM AUTHOR]

A novel method for the creation of machine controllers autonomously is presented. The method is based on reward values which can represent the internal state of the machine or the rating of its task performance. The method consists of a biologically sound constructive neural network model with a minimum number of neurons and no connections initially. New connections and neurons are added when feedback is fed into the neural network using positive and negative reward values. This way the topology and the level of connectivity of the network are kept to a minimum. The method will be applied to a controller for an autonomous mobile robot. [ABSTRACT FROM AUTHOR]

Microarrays are being currently used for the expression levels of thousands of genes simultaneously. They present new analytical challenges because they have a very high input dimension and a very low sample size. It is highly complex to analyse multi-dimensional data with complex geometry and to identify low-dimensional "principal objects" that relate to the optimal projection while losing the least amount of information. Several methods have been proposed for dimensionality reduction of microarray data. Some of these methods include principal component analysis and principal manifolds. This article presents a comparison study of the performance of the linear principal component analysis and the non linear local tangent space alignment principal manifold methods on such a problem. Two microarray data sets will be used in this study. A classification model will be created using fully dimensional and dimensionality reduced data sets. To measure the amount of information lost with the two dimensionality reduction methods, the level of performance of each of the methods will be measured in terms of level of generalisation obtained by the classification models on previously unseen data sets. These results will be compared with the ones obtained using the fully dimensional data sets. [ABSTRACT FROM AUTHOR]

The notion of linear separability is widely used in machine learning research. Learning algorithms that use this concept to learn include neural networks (Single Layer Perceptron and Recursive Deterministic Perceptron), and kernel machines (Support Vector Machines). Several algorithms for testing linear separability exist. Some of these methods are computationally intense. Also, several of them will converge if the classes are linearly separable, but will fail to converge otherwise. A fast and efficient test for non linear separability is proposed which can be used to pretest classification data sets for non linear separability thus avoiding expensive computations. This test is based on the convex hull separability method but does not require the computation of the convex hull. [ABSTRACT FROM AUTHOR]

While artificial neural networks (ANNs) are undoubtedly powerful classifiers their results are sometimes treated with suspicion. This is because their decisions are not open to inspection - the knowledge they contain is hidden. In this paper we describe a method for extracting and representing the knowledge within an ANN. Mappings between inputs and output classifications are stored in a table and, for each classification, Structured Query Language (SQL) queries are evolved using a genetic algorithm. Each evolved query is a simple, human-readable representation of the knowledge used by the ANN to decide on the classification based on the inputs. This method can also be used to show how the knowledge within an ANN develops as it is trained, and can help to identify problems that are particularly hard, or easy, for ANNs to classify. [ABSTRACT FROM AUTHOR]

The extraction of knowledge from trained neural networks provides a way for explaining the functioning of a neural network. This is important for artificial networks to gain a wider degree of acceptance. An increasing amount of research has been carried out to develop mechanisms, procedures and techniques for extracting knowledge from trained neural networks. This publication presents some of the current research trends on extracting knowledge from trained neural networks. [ABSTRACT FROM AUTHOR]