Another look at statistical learning theory and regularization.

The paper reviews and highlights distinctions between function-approximation (FA) and VC theory and methodology, mainly within the setting of regression problems and a squared-error loss function, and illustrates empirically the differences between the two when data is sparse and/or input distribution is non-uniform. In FA theory, the goal is to estimate an unknown true dependency (or 'target' function) in regression problems, or posterior probability P(y/x) in classification problems. In VC theory, the goal is to 'imitate' unknown target function, in the sense of minimization of prediction risk or good 'generalization'. That is, the result of VC learning depends on (unknown) input distribution, while that of FA does not. This distinction is important because regularization theory originally introduced under clearly stated FA setting [Tikhonov, N. (1963). On solving ill-posed problem and method of regularization. Doklady Akademii Nauk USSR, 153, 501-504; Tikhonov, N., & V. Y. Arsenin (1977). Solution of ill-posed problems. Washington, DC: W. H. Winston], has been later used under risk-minimization or VC setting. More recently, several authors [Evgeniou, T., Pontil, M., & Poggio, T. (2000). Regularization networks and support vector machines. Advances in Computational Mathematics, 13, 1-50; Hastie, T., Tibshirani, R., & Friedman, J. (2001). The elements of statistical learning: Data mining, inference and prediction. Springer; Poggio, T. and Smale, S., (2003). The mathematics of learning: Dealing with data. Notices of the AMS, 50 (5), 537-544] applied constructive methodology based on regularization framework to learning dependencies from data (under VC-theoretical setting). However, such regularization-based learning is usually presented as a purely constructive methodology (with no clearly stated problem setting). This paper compares FA/regularization and VC/risk minimization methodologies in terms of underlying theoretical assumptions. The control of model complexity, using regularization and using the concept of margin in SVMs, is contrasted in the FA and VC formulations.