Bayesian Interpolation-英文文献.pdf

Bayesian Interp olation David J.C. MacKay Computation and Neural Systems California Institute of Technology 139{74 Pasadena CA 91125 mackay@hope.caltech.edu May 18, 1991 Submitted to Neural Computation as a review pap er Abstract mo del, using radial basis functions, using p olynomials, or using feedforward neural networks. At the rst level of infer- Although Bayesian analysis has b een in use since Laplace, ence, we take each mo del individually and nd the b est t the Bayesian metho d of model{comparison has only recently interpolant for that mo del. At the second level of inference been developed in depth. we want to rank the alternative mo dels and state for our par- In this pap er, the Bayesian approach to regularisation and ticular data set that, for example, splines are probably the model{comparison is demonstrated by studying the inference b est interp olation mo del, or if the interpolant is mo delled as problem of interpolating noisy data. The concepts and meth- a p olynomial, it should probably b e a cubic. ods describ ed are quite general and can b e applied to many Bayesian metho ds are able consistently and quantitatively other problems. to solve b oth these inference tasks. There is a p opular myth Regularising constants