Tom Mitchell states that the goal of his text Machine Learning is to present the key algorithms and theory that form the core of machine learning. Not only has Mitchell suc-ceeded in his primary goal, but he has accomplished a number of other important goals.

MetLife processes over 260,000 life insurance applications a year. Underwriting of these applications is labor intensive. Automation is difficult because the applications include many free-form text fields. MetLife's intelligent text analyzer (MITA) uses the information-extraction technique of natural language processing to structure the extensive textual fields on a life insurance application. Knowledge engineering, with the help of underwriters as domain experts, was performed to elicit significant concepts for both medical and occupational textual fields. A corpus of 20,000 life insurance applications provided the syntactical and semantic patterns in which these underwriting concepts occur. These patterns, in conjunction with the concepts, formed the frameworks for information extraction. Extension of the information-extraction work developed by Wendy Lehnert was used to populate these frameworks with classes obtained from the systematized nomenclature of human and veterinary medicine and the Dictionary of Occupational Titles ontologies. These structured frameworks can then be analyzed by conventional knowledge-based systems. MITA is currently processing 20,000 life insurance applications a month. Eighty-nine percent of the textual fields processed by MITA exceed the established confidence-level threshold and are potentially available for further analysis by domain-specific analyzers.

We present a system, CREWS_NS, that is used in the long-term scheduling of drivers and guards for the Dutch Railways. CREWS_NS schedules the work of about 5000 people. CREWS_NS is built on top of CREWS, a scheduling tool for speeding the development of scheduling applications. CREWS heavily relies on the use of AI techniques and has been built as a white-box system, in the sense that the planner can perceive what is going on, can interact with the system by proposing alternatives or querying decisions, and can adapt the behavior of the system to changing circumstances. Scheduling can be done in automatic, semiautomatic, or manual mode. CREWS has mechanisms for dealing with the constant changes that occur in input data, can identify the consequences of the change, and guides the planner in accommodating the changes in the already built schedules (rescheduling).

In this paper we consider the problem of `theory patching', in which we are given a domain theory, some of whose components are indicated to be possibly flawed, and a set of labeled training examples for the domain concept. The theory patching problem is to revise only the indicated components of the theory, such that the resulting theory correctly classifies all the training examples. Theory patching is thus a type of theory revision in which revisions are made to individual components of the theory. Our concern in this paper is to determine for which classes of logical domain theories the theory patching problem is tractable. We consider both propositional and first-order domain theories, and show that the theory patching problem is equivalent to that of determining what information contained in a theory is `stable' regardless of what revisions might be performed to the theory. We show that determining stability is tractable if the input theory satisfies two conditions: that revisions to each theory component have monotonic effects on the classification of examples, and that theory components act independently in the classification of examples in the theory. We also show how the concepts introduced can be used to determine the soundness and completeness of particular theory patching algorithms.

The following investigates the use of single-neuron learning algorithms to improve the performance of text-retrieval systems that accept natural-language queries. A retrieval process is explained that transforms the natural-language query into the query syntax of a real retrieval system: the initial query is expanded using statistical and learning techniques and is then used for document ranking and binary classification. The results of experiments suggest that Kivinen and Warmuth's Exponentiated Gradient Descent learning algorithm works significantly better than previous approaches. 1 Introduction The following work explores two learning algorithms - Least Mean Squared (LMS) [1] and Exponentiated Gradient Descent (EG) [2] - in the context of text-based Information Retrieval (IR) systems. The experiments presented in [3] use connectionist to improve the retrieval of relevant documents from a largelearning models collection of text. Previous the area employs various techniques for improving retrieval [6, 7, 14].

Two dimensional image motion detection neural networks have been implemented using a general purpose analog neural computer. The neural circuits perform spatiotemporal feature extraction based on the cortical motion detection model of Adelson and Bergen. The neural computer provides the neurons, synapses and synaptic time-constants required to realize the model in VLSI hardware. Results show that visual motion estimation can be implemented with simple sum-andthreshold neural hardware with temporal computational capabilities. The neural circuits compute general 20 visual motion in real-time.

It has been suggested that long-range intrinsic connections in striate cortex may play a role in contour extraction (Gilbert et aI., 1996). A number of recent physiological and psychophysical studies have examined the possible role of long range connections in the modulation of contrast detection thresholds (Polat and Sagi, 1993,1994; Kapadia et aI., 1995; Kovacs and Julesz, 1994) and various pre-attentive detection tasks (Kovacs and Julesz, 1993; Field et aI., 1993). We have developed a network architecture based on the anatomical connectivity of striate cortex, as well as the temporal dynamics of neuronal processing, that is able to reproduce the observed experimental results. The network has been tested on real images and has applications in terms of identifying salient contours in automatic image processing systems. 1 INTRODUCTION Vision is an active process, and one of the earliest, preattentive actions in visual processing is the identification of the salient contours in a scene. We propose that this process depends upon two properties of striate cortex: the pattern of horizontal connections between orientation columns, and temporal synchronization of cell responses. In particular, we propose that perceptual salience is directly related to the degree of cell synchronization. We present results of network simulations that account for recent physiological and psychophysical "pop-out" experiments, and which successfully extract salient contours from real images.

To obtain classification systems with both good generalization performance and efficiency in space and time, we propose a learning method based on combinations of weak classifiers, where weak classifiers are linear classifiers (perceptrons) which can do a little better than making random guesses. A randomized algorithm is proposed to find the weak classifiers. They· are then combined through a majority vote. As demonstrated through systematic experiments, the method developed is able to obtain combinations of weak classifiers with good generalization performance and a fast training time on a variety of test problems and real applications.

High frequency foreign exchange data can be decomposed into three components: the inventory effect component, the surprise infonnation (news) component and the regular infonnation component. The presence of the inventory effect and news can make analysis of trends due to the diffusion of infonnation (regular information component) difficult. We propose a neural-net-based, independent component analysis to separate high frequency foreign exchange data into these three components. Our empirical results show that our proposed multi-effect decomposition can reveal the intrinsic price behavior.

We compare different methods to combine predictions from neural networks trained on different bootstrap samples of a regression problem. One of these methods, introduced in [6] and which we here call balancing, is based on the analysis of the ensemble generalization error into an ambiguity term and a term incorporating generalization performances of individual networks. We show how to estimate these individual errors from the residuals on validation patterns. Weighting factors for the different networks follow from a quadratic programming problem. On a real-world problem concerning the prediction of sales figures and on the well-known Boston housing data set, balancing clearly outperforms other recently proposed alternatives as bagging [1] and bumping [8]. 1 EARLY STOPPING AND BOOTSTRAPPING Stopped training is a popular strategy to prevent overfitting in neural networks.