作者: Judea Pearl
DOI: 10.1109/TPAMI.1979.4766943
关键词:
摘要: This paper extends the notions of capacity and distribution-free error estimation to nonlinear Boolean classifiers on patterns with binary-valued features. We establish quantitative relationships between dimensionality feature vectors (d), combinational complexity decision rule (c), number samples in training set (n), classification performance resulting classifier. Our results state that discriminating is given by product dc, probability ambiguous generalization asymptotically (n/dc-1)-1 0(log d)/d) for large d, n=0(dc). In addition we show if a fraction ? misclassified then (?) subsequent satisfies P(|?-?| ?) m=<2.773 exp (dc-e2n/8) all distributions, regardless how classifier was discovered.