Topic: Mixing of the data
All greetings! There are some dial-ups of the heterogeneous data for training there is nobody the qualifier. For example, the qualifier of persons also is some bases of persons with different characteristics of photos (lighting, noise, photo resolution....) . Also there are some test bases on which quality of the qualifier (their characteristic is checked can differ from training). Let training procedure is fixed: casually selected given amount of the data from one specific basis undertakes (for example 10000 photos), are trained the qualifier on the given algorithm with the fixed parameters. After that it is tested on each of test bases, the dial-up of values of accuracy turns out. Then the same becomes for other learning basis, for thirds etc., bases big, on 10000 photos in everyone is. It is obvious that depending on character of learning basis of value of accuracy on test will walk somehow - somewhere to be refined, somewhere to worsen. Attention, a question: how to select proportions in which it is necessary to mix the learning data from different bases that on the received 10000 photos the qualifier it was trained in the best way (in some sense "averaged" on all test cases)? I.e. How to use the information from accuracy of each of the trained qualifiers on all tests for optimal selection of composition of a learning COMPOUND. I suspect that the question is researched. Stick into links or tell.