Department of Electrical and Computer Engineering University of Queensland St. Lucia Qld 4072 Australia Abstract We define a Gamma multi-layer perceptron (MLP) as an MLP with the usual synaptic weights replaced by gamma filters (as proposed byde Vries and Principe (de Vries and Principe, 1992)) and associated gain terms throughout all layers. We derive gradient descent update equations and apply the model to the recognition of speech phonemes. We find that both the inclusion of gamma filters in all layers, and the inclusion of synaptic gains, improves the performance of the Gamma MLP. We compare the Gamma MLP with TDNN, Back-Tsoi FIR MLP, and Back-Tsoi I1R MLP architectures, and a local approximation scheme. We find that the Gamma MLP results in an substantial reduction in error rates. 1 INTRODUCTION 1.1 THE GAMMA FILTER Infinite Impulse Response (I1R) filters have a significant advantage over Finite Impulse Response(FIR) filters in signal processing: the length of the impulse response is uncoupled from the number of filter parameters.

Central to the performance improvement of a committee relative to individual networks is the error correlation between networks in the committee. We investigated methods of achieving error independence betweenthe networks by training the networks with different resampling sets from the original training set. The methods were tested on the sinwave artificial task and the real-world problems of hepatoma (liver cancer) and breast cancer diagnoses.

We present a neural network-based face detection system. A retinally connected neural network examines small windows of an image, and decides whether each window contains a face. The system arbitrates between multiple networks to improve performance over a single network. We use a bootstrap algorithm for training, which adds false detections into the training set as training progresses. This eliminates the difficult task of manually selecting non-face training examples, which must be chosen to span the entire space of non-face images.

A statistical theory for overtraining is proposed. The analysis treats realizable stochastic neural networks, trained with Kullback Leibler loss in the asymptotic case. It is shown that the asymptotic gain in the generalization error is small if we perform early stopping, evenif we have access to the optimal stopping time. Considering cross-validation stopping we answer the question: In what ratio the examples should be divided into training and testing sets in order toobtain the optimum performance. In the non-asymptotic region cross-validated early stopping always decreases the generalization error.Our large scale simulations done on a CM5 are in nice agreement with our analytical findings.

We work in a setting in which we must choose the right number of parameters for a hypothesis function in response to a finite training sample, with the goal of minimizing the resulting generalization error. There is a large and interesting literature on cross validation methods, which often emphasizes asymptotic statistical properties, or the exact calculation of the generalization error for simple models. Our approach here is somewhat different, and is primariIy inspired by two sources. The first is the work of Barron and Cover [2], who introduced the idea of bounding the error of a model selection method (in their case, the Minimum Description Length Principle) in terms of a quantity known as the index of resolvability. The second is the work of Vapnik [5], who provided extremely powerful and general tools for uniformly bounding the deviations between training and generalization errors. We combine these methods to give a new and general analysis of cross validation performance. Inthe first and more formal part of the paper, we give a rigorous bound on the error of cross validation in terms of two parameters of the underlying model selection problem: the approximation rate and the estimation rate. In the second and more experimental part of the paper, we investigate the implications of our bound for choosing'Y, the fraction of data withheld for testing in cross validation. The most interesting aspect of this analysis is the identification of several qualitative properties of the optimal'Y that appear to be invariant over a wide class of model selection problems: - When the target function complexity is small compared to the sample size, the performance of cross validation is relatively insensitive to the choice of'Y.

This paper presents new experimental evidence against the utility of Occam's razor. A~systematic procedure is presented for post-processing decision trees produced by C4.5. This procedure was derived by rejecting Occam's razor and instead attending to the assumption that similar objects are likely to belong to the same class. It increases a decision tree's complexity without altering the performance of that tree on the training data from which it is inferred. The resulting more complex decision trees are demonstrated to have, on average, for a variety of common learning tasks, higher predictive accuracy than the less complex original decision trees. This result raises considerable doubt about the utility of Occam's razor as it is commonly applied in modern machine learning.

A reported weakness of C4.5 in domains with continuous attributes is addressed by modifying the formation and evaluation of tests on continuous attributes. An MDL-inspired penalty is applied to such tests, eliminating some of them from consideration and altering the relative desirability of all tests. Empirical trials show that the modifications lead to smaller decision trees with higher predictive accuracies. Results also confirm that a new version of C4.5 incorporating these changes is superior to recent approaches that use global discretization and that construct small trees with multi-interval splits.

MLP networks provided slightly better risk prediction than conventional logistic regression when used to predict the risk of death, stroke, and renal failure on 1257 patients who underwent coronary artery bypass operations. Bootstrap sampling was required to compare approaches and regularization provided by early stopping was an important component of improved performance. A simplified approach to generating confidence intervals for MLP risk predictions using an auxiliary "confidence MLP" was also developed. The confidence MLP is trained to reproduce the confidence bounds that were generated during training by 50 MLP networks trained using bootstrap samples. Current research is validating these results using larger data sets, exploring approaches to detect outlier patients who are so different from any training patient that accurate risk prediction is suspect, developing approaches to explaining which input features are important for an individual patient, and determining why MLP networks provide improved performance.

Instead of "ground truth" one may only have the subjective opinion(s) of one or more experts. For example, medical data or image data may be collected off-line and some time later a set of experts analyze the data and produce a set of class labels. The central problem is that of trying to infer the "ground truth" given the noisy subjective estimates of the experts. When one wishes to apply a supervised learning algorithm to the data, the problem is primarily twofold: (i) how to evaluate the relative performance of experts and algorithms, and (ii) how to train a pattern recognition system in the absence of absolute ground truth. In this paper we focus on problem (i), namely the performance evaluation issue, and in particular we discuss the application of a particular modelling technique to the problem of counting volcanoes on the surface of Venus.

An analogue VLSI neural network has been designed and tested to perform cardiac morphology classification tasks. Analogue techniques were chosen to meet the strict power and area requirements of an Implantable Cardioverter Defibrillator (ICD) system. The robustness of the neural network architecture reduces the impact of noise, drift and offsets inherent in analogue approaches. The network is a 10:6:3 multi-layer percept ron with on chip digital weight storage, a bucket brigade input to feed the Intracardiac Electrogram (ICEG) to the network and has a winner take all circuit at the output. The network was trained in loop and included a commercial ICD in the signal processing path. The system has successfully distinguished arrhythmia for different patients with better than 90% true positive and true negative detections for dangerous rhythms which cannot be detected by present ICDs. The chip was implemented in 1.2um CMOS and consumes less than 200n W maximum average power in an area of 2.2 x 2.2mm2.