Where should a researcher conduct experiments to provide training data for a multilayer perceptron? This question is investigated, and a statistical method for selecting optimal experimental design points for multiple output multilayer perceptrons is introduced. Multiple class discrimination problems are examined using a framework in which the multilayer perceptron is viewed as a multivariate nonlinear regression model. Following a Bayesian formulation for the case where the variance-covariance matrix of the responses is unknown, a selection criterion is developed. This criterion is based on the volume of the joint confidence ellipsoid for the weights in a multilayer perceptron. An example is used to demonstrate the superiority of optimally selected design points over randomly chosen points, as well as points chosen in a grid pattern. Simplification of the basic criterion is offered through the use of Hadamard matrices to produce uncorrelated outputs.

A new model-based vision (MBV) algorithm is developed to find regions of interest (ROI's) corresponding to masses in digitized mammograms and to classify the masses as malignant/benign. The MBV algorithm is comprised of five modules to structurally identify suspicious ROI's, eliminate false positives, and classify the remaining as malignant or benign. The focus of attention module uses a difference of Gaussians (DoG) filter to highlight suspicious regions in the mammogram. The index module uses tests to reduce the number of nonmalignant regions from 8.39 to 2.36 per full breast image. Size, shape, contrast, and Laws texture features are used to develop the prediction module's mass models. Derivative-based feature saliency techniques are used to determine the best features for classification. Nine features are chosen to define the malignant/benign models. The feature extraction module obtains these features from all suspicious ROI's. The matching module classifies the regions using a multilayer perceptron neural network architecture to obtain an overall classification accuracy of 100% for the segmented malignant masses with a false-positive rate of 1.8 per full breast image. This system has a sensitivity of 92% for locating malignant ROI's. The database contains 272 images (12 b, 100 microm) with 36 malignant and 53 benign mass images. The results demonstrate that the MBV approach provides a structured order of integrating complex stages into a system for radiologists.

Automatic speaker recognition systems are maturing and databases have been designed to specifically compare algorithms and results to target error rates. The LDC YOHO speaker verification database was designed to test error rates at the 1% false rejection and 0.1% false acceptance level. This work examines the use of speaker-dependent (SD) monophone models to meet these requirements. By representing each speaker with 22 monophones, both closed-set speaker identification and global-threshold verification was performed. Using four combination lock phrases, speaker identification error rates are obtained at 0.19% for males and 0.31% for females. By defining a test hypothesis, a critical error analysis for speaker verification is developed and new results reported for YOHO. A new Bhattacharyya distance is developed for cohort selection. This method, based on the second order statistics of the enrolment Viterbi log-likelihoods, determines the optimal cohorts and achieves an equal error rate of 0.282%.

This paper presents the results of experimentation with a simple ultrasonic lip motion detector or "Ultrasonic Mike" in automatic speech recognition. The device is tested in a speaker dependent isolated word recognition task with a vocabulary consisting of the spoken digits from zero to nine. The "Ultrasonic Mike" is used as input to an automatic lip reader. The automatic lip reader uses template matching and dynamic time warping to determine the best candidate for a given test utterance. The device is first tested as a stand alone automatic lip reader achieving accuracy as high as 89%. Next the automatic lip reader is combined with a conventional automatic speech recognizer. Classifier fusion is based on a pseudo probability mass function derived from the dynamic time warping distances. The combined system is tested with various levels of acoustic noise added. In a typical example, at 0 dB, the acoustic recognizer's accuracy was 78%, the lip reader accuracy was at 69%, but the combined accuracy was 93%. This experiment demonstrates that this simple ultrasonic lip motion detector, that has an output data rate 12500 times less than a typical video camera, can improve automatic speech recognition in noisy environments. This experiment also demonstrates an effective classifier fusion algorithm based on dynamic time warping distances.

The relationship between backpropagation and extended Kalman filtering for training multilayer perceptrons is examined. These two techniques are compared theoretically and empirically using sensor imagery. Backpropagation is a technique from neural networks for assigning weights in a multilayer perceptron. An extended Kalman filter can also be used for this purpose. A brief review of the multilayer perceptron and these two training methods is provided. Then, it is shown that backpropagation is a degenerate form of the extended Kalman filter. The training rules are compared in two examples: an image classification problem using laser radar Doppler imagery and a target detection problem using absolute range images. In both examples, the backpropagation training algorithm is shown to be three orders of magnitude less costly than the extended Kalman filter algorithm in terms of a number of floating-point operations.>

The problem of fusing information from multiple sensors to perform pattern recognition is addressed. A technique is proposed for performing target/nontarget discrimination using information from absolute range and forward looking infrared (FLIR) sensors. A multilayer perceptron is used to perform the feature-level fusion and is compared to ak-nearest-neighbor classifier as well as a nonparametric Bayesian classifier using Parzen windows. All three classifiers show statistically significant improvement from the fusion process; however, the multilayer perceptron uses far fewer free parameters, which should improve generalization capability. Also, the multilayer perceptron requires much less computation than the traditional classifiers

The multilayer perceptron, when trained as a classifier using backpropagation, is shown to approximate the Bayes optimal discriminant function. The result is demonstrated for both the two-class problem and multiple classes. It is shown that the outputs of the multilayer perceptron approximate the a posteriori probability functions of the classes being trained. The proof applies to any number of layers and any type of unit activation function, linear or nonlinear.