Part of Advances in Neural Information Processing Systems 14 (NIPS 2001)
John Langford, Rich Caruana
We present a new approach to bounding the true error rate of a continuous valued classiﬁer based upon PAC-Bayes bounds. The method ﬁrst con- structs a distribution over classiﬁers by determining how sensitive each parameter in the model is to noise. The true error rate of the stochastic classiﬁer found with the sensitivity analysis can then be tightly bounded using a PAC-Bayes bound. In this paper we demonstrate the method on artiﬁcial neural networks with results of a order of magnitude im- provement vs. the best deterministic neural net bounds.