A family of cascade NN-like classifiers

We propose a family of two-stage classifiers with one common feature: the first stage classifier is an ensemble of classifiers working in parallel and the prediction in this stage is considered confident (and thus final) if and only if all the component classifiers return the same answer. If there is no consensus, the sample is sent to the second stage with a costlier classifier. An attractive accuracy-speed tradeoff of the presented algorithms is experimentally confirmed.