Abstract. One of the surprising recurring phenomena observed in experiments with boosting is that the test error of the generated classifier usually does not increase as its size becomes very large, and often is observed to decrease even after the training error reaches zero. In this paper, we show that this phenomenon is related to the distribution of margins of the training examples with respect to the generated voting classification rule, where the margin of an example is simply the difference between the number of correct votes and the maximum number of votes received by any incorrect label. We show that techniques used in the analysis of Vapnik’s support vector classifiers and of neural networks with small weights can be applied to voting methods to relate the margin distribution to the test error. We also show theoretically and experimentally that boosting is especially effective at increasing the margins of the training examples. Finally, we compare our explanation to those based on the bias-variance decomposition. 1

Sample complexity results from computational learning theory, when applied to neural network learning for pattern classification problems, suggest that for good generalization performance the number of training examples should grow at least linearly with the number of adjustable parameters in the network. Results in this paper show that if a large neural network is used for a pattern classification problem and the learning algorithm finds a network with small weights that has small squared error on the training patterns, then the generalization performance depends on the size of the weights rather than the number of weights. For example, consider a two-layer feedforward network of sigmoid units, in which the sum of the magnitudes of the weights associated with each unit is bounded by A and the input dimension is n. We show that the misclassification probability is no more than a certain error estimate (that is related to squared error on the training set) plus A³ p (log n)=m (ignori...

We propose new bounds on the error of learning algorithms in terms of a data-dependent notion of complexity. The estimates we establish give optimal rates and are based on a local and empirical version of Rademacher averages, in the sense that the Rademacher averages are computed from the data, on a subset of functions with small empirical error. We present some applications to classification and prediction with convex function classes, and with kernel classes in particular.

Many of the classification algorithms developed in the machine learning literature, including the support vector machine and boosting, can be viewed as minimum contrast methods that minimize a convex surrogate of the 0-1 loss function. The convexity makes these algorithms computationally efficient. The use of a surrogate, however, has statistical consequences that must be balanced against the computational virtues of convexity. To study these issues, we provide a general quantitative relationship between the risk as assessed using the 0-1 loss and the risk as assessed using any nonnegative surrogate loss function. We show that this relationship gives nontrivial upper bounds on excess risk under the weakest possible condition on the loss function: that it satisfy a pointwise form of Fisher consistency for classification. The relationship is based on a simple variational transformation of the loss function that is easy to compute in many applications. We also present a refined version of this result in the case of low noise. Finally, we

Gaussian mixtures (or so-called radial basis function networks) for density estimation provide a natural counterpart to sigmoidal neural networks for function tting and approximation. In both cases, it is possible to give simple expressions for the iterative improvement of performance as components of the network are introduced one at a time. In particular, for mixture density estimation we show that a k-component mixture estimated by maximum likelihood (or by an iterative likelihood improvement that we introduce) achieves log-likelihood within order 1=k of the log-likelihood achievable by any convex combination. Consequences for approximation and estimation using Kullback-Leibler risk are also given. A Minimum Description Length principle selects the optimal number of components k that minimizes the risk bound. 1 Introduction In density estimation, Gaussian mixtures provide exible-basis representations for densities that can be used to model heterogeneous data in high dimensions....

Recently, sample complexity bounds have been derived for problems involving linear functions such as neural networks and support vector machines. In many of these theoretical studies, the concept of covering numbers played an important role. It is thus useful to study covering numbers for linear function classes. In this paper, we investigate two closely related methods to derive upper bounds on these covering numbers. The first method, already employed in some earlier studies, relies on the so-called Maurey's lemma; the second method uses techniques from the mistake bound framework in online learning. We compare results from these two methods, as well as their consequences in some learning formulations.

Abstract Boosting is one of the most significant advances in machine learning for classification and regression. In its original and computationally flexible version, boosting seeks to minimize empirically a loss function in a greedy fashion. The resulted estimator takes an additive function form and is built iteratively by applying a base estimator (or learner) to updated samples depending on the previous iterations. An unusual regularization technique, early stopping, is employed based on CV or a test set. This paper studies numerical convergence, consistency, and statistical rates of convergence of boosting with early stopping, when it is carried out over the linear span of a family of basis functions. For general loss functions, we prove the convergence of boosting's greedy optimization to the infinimum of the loss function over the linear span. Using the numerical convergence result, we find early stopping strategies under which boosting is shown to be consistent based on iid samples, and we obtain bounds on the rates of convergence for boosting estimators. Simulation studies are also presented to illustrate the relevance of our theoretical results for providing insights to practical aspects of boosting. As a side product, these results also reveal the importance of restricting the greedy search step sizes, as known in practice through the works of Friedman and others. Moreover, our results lead to a rigorous proof that for a linearly separable problem, AdaBoost with ffl! 0 stepsize becomes an L1-margin maximizer when left to run to convergence. 1 Introduction In this paper we consider boosting algorithms for classification and regression. These algorithms present one of the major progresses in machine learning. In their original version, the computational aspect is explicitly specified as part of the estimator/algorithm. That is, the empirical minimization of an appropriate loss function is carried out in a greedy fashion, which means that at each step, a basis function that leads to the largest reduction of empirical risk is added into the estimator. This specification distinguishes boosting from other statistical procedures which are defined by an empirical minimization of a loss function without the numerical optimization details.

Abstract. We consider the problem of one-step ahead prediction for time series generated by an underlying stationary stochastic process obeying the condition of absolute regularity, describing the mixing nature of process. We make use of recent results from the theory of empirical processes, and adapt the uniform convergence framework of Vapnik and Chervonenkis to the problem of time series prediction, obtaining finite sample bounds. Furthermore, by allowing both the model complexity and memory size to be adaptively determined by the data, we derive nonparametric rates of convergence through an extension of the method of structural risk minimization suggested by Vapnik. All our results are derived for general L p error measures, and apply to both exponentially and algebraically mixing processes.

In many classification procedures, the classification function is obtained (or trained) by minimizing a certain empirical risk on the training sample. The classification is then based on the sign of the classification function. In recent years, there have been a host of classification methods proposed in machine learning that use di#erent margin-based loss functions in the training. Examples include the AdaBoost procedure, the support vector machine, and many variants of them. The margin-based loss functions used in these procedures are usually motivated as upper bounds of the misclassification loss, but this can not explain the statistical properties of the classification procedures. We consider the margin-based loss functions from a statistical point of view. We first show that under general conditions, margin-based loss functions are Fisher consistent for classification. That is, the population minimizer of the loss function leads to the Bayes optimal rule of classification. In particular, almost all margin-based loss functions that have appeared in the literature are Fisher consistent. We then study margin-based loss functions in the method of sieves and the method of regularization. We show that the Fisher consistency of margin-based loss functions often leads to consistency and rate of convergence (to the Bayes optimal risk) results under general conditions. The common notion of margin-based loss functions as upper bounds of the misclassification loss is formalized and investigated. It is shown that the hinge loss is the tightest convex upper bound of the misclassification loss. Simulations are carried out to compare some commonly used margin-based loss functions.

Introduction Previous negative results for learning two-layer neural network classifiers show that it is difficult to find a network that correctly classifies all examples in a training set. However, for learning to a particular accuracy it is only necessary to approximately solve this problem, that is, to find a network that correctly classifies most examples in a training set. In this paper, we show that this approximation problem is hard for several neural network classes. The hardness of PAC style learning is a very natural question that has been addressed from a variety of viewpoints. The strongest non-learnability conclusions are those stating that no matter what type of algorithm a learner may use, as long as its computational resources are limited, it would not be able to predict a previously unseen label (with probability significantly better than that of a random guess). Such results have been derived by noticing that, in some precise sense, learning