Abstract-This paper investigates the application of the mutual infor “ criterion to evaluate a set of candidate features and to select an informative subset to be used as input data for a neural network classifier. Because the mutual information measures arbitrary dependencies between random variables, it is suitable for assessing the “information content ” of features in complex classification tasks, where methods bases on linear relations (like the correlation) are prone to mistakes. The fact that the mutual information is independent of the coordinates chosen permits a robust estimation. Nonetheless, the use of the mutual information for tasks characterized by high input dimensionality requires suitable approximations because of the prohibitive demands on computation and samples. An algorithm is proposed that is based on a “greedy ” selection of the features and that takes both the mutual information with respect to the output class and with respect to the already-selected features into account. Finally the results of a series of experiments are discussed. Index Terms-Feature extraction, neural network pruning, di-mensionality reduction, mutual information, supervised learning,

Classification methods from statistical pattern recognition, neural nets, and machine learning were applied to four real-world data sets. Each of these data sets has been previously analyzed and reported in the statistical, medical, or machine learning literature. The data sets are characterized by statisucal uncertainty; there is no completely accurate solution to these problems. Training and testing or resampling techniques are used to estimate the true error rates of the classification methods. Detailed attention is given to the analysis of performance of the neural nets using back propagation. For these problems, which have relatively few hypotheses and features, the machine learning procedures for rule induction or tree induction clearly performed best. 1

Abstract | We survey learning algorithms for recurrent neural networks with hidden units, and put the various techniques into a common framework. We discuss xedpoint learning algorithms, namely recurrent backpropagation and deterministic Boltzmann Machines, and non- xedpoint algorithms, namely backpropagation through time, Elman's history cuto, and Jordan's output feedback architecture. Forward propagation, an online technique that uses adjoint equations, and variations thereof, are also discussed. In many cases, the uni ed presentation leads to generalizations of various sorts. We discuss advantages and disadvantages of temporally continuous neural networks in contrast to clocked ones, continue with some \tricks of the trade" for training, using, and simulating continuous time and recurrent neural networks. We present somesimulations, and at the end, address issues of computational complexity and learning speed.

Nearest-neighbour interpolation algorithms have many useful properties for applications to learning, but they often exhibit poor generalization. In this paper, it is shown that much better generalization can be obtained by using a variable interpolation kernel in combination with conjugate gradient optimization of the similarity metric and kernel size. The resulting method is called variable-kernel similarity metric (VSM) learning. It has been tested on several standard classification data sets, and on these problems it shows better generalization than back propagation and most other learning methods. An important advantage is that the system can operate as a black box in which no model minimization parameters need to be experimentally set by the user. The number of parameters that must be determined through optimization are orders of magnitude less than for back-propagation or RBF networks, which may indicate that the method better captures the essential degrees of variation in learni...

Many lazy learning algorithms are derivatives of the k-nearest neighbor (k-NN) classifier, which uses a distance function to generate predictions from stored instances. Several studies have shown that k-NN's performance is highly sensitive to the definition of its distance function. Many k-NN variants have been proposed to reduce this sensitivity by parameterizing the distance function with feature weights. However, these variants have not been categorized nor empirically compared. This paper reviews a class of weight-setting methods for lazy learning algorithms. We introduce a framework for distinguishing these methods and empirically compare them. We observed four trends from our experiments and conducted further studies to highlight them. Our results suggest that methods which use performance feedback to assign weight settings demonstrated three advantages over other methods: they require less pre-processing, perform better in the presence of interacting features, and generally require less training data to learn good settings. We also found that continuous weighting methods tend to outperform feature selection algorithms for tasks where some features are useful but less important than others.

Fisher's linear discriminant analysis is a valuable tool for multigroup classification. With a large number of predictors, one can nd a reduced number of discriminant coordinate functions that are "optimal" for separating the groups. With two such functions one can produce a classification map that partitions the reduced space into regions that are identified with group membership, and the decision boundaries are linear. This paper is about richer nonlinear classification schemes. Linear discriminant analysis is equivalent to multi-response linear regression using optimal scorings to represent the groups. We obtain nonparametric versions of discriminant analysis by replacing linear regression by any nonparametric regression method. In this way, any multi-response regression technique (such as MARS or neural networks) can be post-processed to improve their classification performence.

To Mom, Dad, and Susan, for their support and encouragement.

We derive a mean field algorithm for binary classification with Gaussian processes which is based on the TAP approach originally proposed in Statistical Physics of disordered systems. The theory also yields an approximate leave-one-out estimator for the generalization error which is computed with no extra computational cost. We show that from the TAP approach, it is possible to derive both a simpler `naive' mean field theory and support vector machines (SVM) as limiting cases. For both mean field algorithms and support vectors machines, simulation results for three small benchmark data sets are presented. They show 1. that one may get state of the art performance by using the leave-one-out estimator for model selection and 2. the built-in leave-one-out estimators are extremely precise when compared to the exact leave-one-out estimate. The latter result is a taken as a strong support for the internal consistency of the mean field approach. 1 1

Novelty Detection techniques are conceptlearning methods that proceed by recognizing positive instances of a concept rather than differentiating between its positive and negative instances. Novelty Detection approaches consequently require very few, if any, negative training instances. This paper presents a particular Novelty Detection approach to classification that uses a Redundancy Compression and Non-Redundancy Differentiation technique based on the [Gluck & Myers, 1993] model of the hippocampus, a part of the brain critically involved in learning and memory. In particular, this approach consists of training an autoencoder to reconstruct positive input instances at the output layer and then using this autoencoder to recognize novel instances. Classification is possible, after training, because positive instances are expected to be reconstructed accurately while negative instances are not. The purpose of this paper is to compare HIPPO, the system that implements this technique, to C4.5 and feedforward neural network classification on several applications. 1

Feature selection is an integral part of most learning algorithms. Due to the existence of irrelevant and redundant attributes, by selecting only the relevant attributes of the data, higher predictive accuracy can be expected from a machine learning method. In this paper, we propose the use of a three-layer feedforward neural network to select those input attributes that are most useful for discriminating classes in a given set of input patterns. A network pruning algorithm is the foundation of the proposed algorithm. By adding a penalty term to the error function of the network, redundant network connections can be distinguished from those relevant ones by their small weights when the network training process has been completed. A simple criterion to remove an attribute based on the accuracy rate of the network is developed. The network is retrained after removal of an attribute, and the selection process is repeated until no attribute meets the criterion for removal. Our ...