Integrated Mean Squared Error (IMSE) is a version of the usual mean squared error criterion, averaged over all possible training sets of a given size. If it could be observed, it could be used to determine optimal network complexity or optimal data subsets forefficient training. We show that two common methods of cross-validating average squared error deliver unbiased estimates of IMSE, converging to IMSE with probability one. We also show that two variants of cross validation measure provide unbiased IMSE-based estimates potentially useful for selecting optimal data subsets. 1 Summary To begin, assume we are given a fixed network architecture. Let zN denote a given set of N training examples.

Thomas G. Dietterich Arris Pharmaceutical Corporation and Oregon State University Corvallis, OR 97331-3202 Ajay N. Jain Arris Pharmaceutical Corporation 385 Oyster Point Blvd., Suite 3 South San Francisco, CA 94080 Richard H. Lathrop and Tomas Lozano-Perez Arris Pharmaceutical Corporation and MIT Artificial Intelligence Laboratory 545 Technology Square Cambridge, MA 02139 Abstract In drug activity prediction (as in handwritten character recognition), thefeatures extracted to describe a training example depend on the pose (location, orientation, etc.) of the example. In handwritten characterrecognition, one of the best techniques for addressing thisproblem is the tangent distance method of Simard, LeCun and Denker (1993). Jain, et al. (1993a; 1993b) introduce a new technique-dynamic reposing-that also addresses this problem. Dynamicreposing iteratively learns a neural network and then reposes the examples in an effort to maximize the predicted output values.New models are trained and new poses computed until models and poses converge. This paper compares dynamic reposing to the tangent distance method on the task of predicting the biological activityof musk compounds.

The theory revision problem is the problem of how best to go about revising a deficient domain theory using information contained in examples that expose inaccuracies. In this paper we present our approach to the theory revision problem for propositional domain theories. The approach described here, called PTR, uses probabilities associated with domain theory elements to numerically track the ``flow'' of proof through the theory. This allows us to measure the precise role of a clause or literal in allowing or preventing a (desired or undesired) derivation for a given example. This information is used to efficiently locate and repair flawed elements of the theory. PTR is proved to converge to a theory which correctly classifies all examples, and shown experimentally to be fast and accurate even for deep theories.

How can artificial neural nets generalize better from fewer examples? In order to generalize successfully, neural network learning methods typically require large training data sets. We introduce a neural network learning method that generalizes rationally from many fewer data points, relying instead on prior knowledge encoded in previously learned neural networks. For example, in robot control learning tasks reported here, previously learned networks that model the effects of robot actions are used to guide subsequent learning of robot control functions. For each observed training example of the target function (e.g. the robot control policy), the learner explains the observed example in terms of its prior knowledge, then analyzes this explanation to infer additional information about the shape, or slope, of the target function. This shape knowledge is used to bias generalization when learning the target function. Results are presented applying this approach to a simulated robot task based on reinforcement learning.

How can artificial neural nets generalize better from fewer examples? In order to generalize successfully, neural network learning methods typically require large training data sets. We introduce a neural network learning method that generalizes rationally from many fewer data points, relying instead on prior knowledge encoded in previously learned neural networks. For example, in robot control learning tasks reported here, previously learned networks that model the effects of robot actions are used to guide subsequent learning of robot control functions. For each observed training example of the target function (e.g. the robot control policy), the learner explains the observed example in terms of its prior knowledge, then analyzes this explanation to infer additional information about the shape, or slope, of the target function. This shape knowledge is used to bias generalization when learning the target function. Results are presented applying this approach to a simulated robot task based on reinforcement learning.

Machine Learning: An Artificial Intelligence Approach contains tutorial overviews and research papers representative of trends in the area of machine learning as viewed from an artificial intelligence perspective. The book is organized into six parts. Part I provides an overview of machine learning and explains why machines should learn. Part II covers important issues affecting the design of learning programs-particularly programs that learn from examples. It also describes inductive learning systems.

Three methods for improving the performance of (gaussian) radial basis function (RBF) networks were tested on the NETtaik task. In RBF, a new example is classified by computing its Euclidean distance to a set of centers chosen by unsupervised methods. The application of supervised learning to learn a non-Euclidean distance metric was found to reduce the error rate of RBF networks, while supervised learning of each center's variance resulted in inferior performance. The best improvement in accuracy was achieved by networks called generalized radial basis function (GRBF) networks. In GRBF, the center locations are determined by supervised learning. After training on 1000 words, RBF classifies 56.5% of letters correct, while GRBF scores 73.4% letters correct (on a separate test set). From these and other experiments, we conclude that supervised learning of center locations can be very important for radial basis function learning.

Three methods for improving the performance of (gaussian) radial basis function (RBF) networks were tested on the NETtaik task. In RBF, a new example is classified by computing its Euclidean distance to a set of centers chosen by unsupervised methods. The application of supervised learning to learn a non-Euclidean distance metric was found to reduce the error rate of RBF networks, while supervised learning of each center's variance resulted in inferior performance. The best improvement in accuracy was achieved by networks called generalized radial basis function (GRBF) networks. In GRBF, the center locations are determined by supervised learning. After training on 1000 words, RBF classifies 56.5% of letters correct, while GRBF scores 73.4% letters correct (on a separate test set). From these and other experiments, we conclude that supervised learning of center locations can be very important for radial basis function learning.

An alternative to the typical technique of selecting training examples independently from a fixed distribution is fonnulated and analyzed, in which the current example is presented repeatedly until the error for that item is reduced to some criterion value,; then, another item is randomly selected. The convergence time can be dramatically increased or decreased by this heuristic, depending on the task, and is very sensitive to the value of .

A method for transforming performance evaluation signals distal both in space and time into proximal signals usable by supervised learning algorithms, presented in [Jordan & Jacobs 90], is examined. A simple observation concerning differentiation through models trained with redundant inputs (as one of their networks is) explains a weakness in the original architecture and suggests a modification: an internal world model that encodes action-space exploration and, crucially, cancels input redundancy to the forward model is added. Learning time on an example task, cartpole balancing, is thereby reduced about 50 to 100 times. 1 INTRODUCTION In many learning control problems, the evaluation used to modify (and thus improve) control may not be available in terms of the controller's output: instead, it may be in terms of a spatial transformation of the controller's output variables (in which case we shall term it as being "distal in space"), or it may be available only several time steps into the future (termed as being "distal in time"). For example, control of a robot arm may be exerted in terms of joint angles, while evaluation may be in terms of the endpoint cartesian coordinates; furthermore, we may only wish to evaluate the endpoint coordinates reached after a certain period of time: the co- ·Current address: Computation and Neural Systems Program, California Institute of Technology, Pasadena CA.