PAC-Bayes Analysis Of Maximum Entropy Classification
; Proceedings of the Twelth International Conference on Artificial Intelligence and Statistics, PMLR 5:480-487, 2009.
We extend and apply the PAC-Bayes theorem to the analysis of maximum entropy learning by considering maximum entropy classification. The theory introduces a multiple sampling technique that controls an effective margin of the bound. We further develop a dual implementation of the convex optimisation that optimises the bound. This algorithm is tested on some simple datasets and the value of the bound compared with the test error.