15:49:26	 From Aurelijus Rinkevicius : Are strata mutually independent? It was not obvious.
15:50:53	 From Miguel Cardenas : GP-based data aumentation is based in a) built a GP model and b) later to sample it?
15:52:27	 From Aurelijus Rinkevicius : OK, thank you.
15:52:30	 From Christian Kragh Jespersen : Could you please talk a bit about how you augment the sparse data area?
15:53:53	 From Miguel Cardenas : ok. Thank you.
15:55:36	 From Marc Chaumont : If I understand well (?) the classification is done by
1) select the good strate classifier (how?) ?
2) then use the classifier associated to the strate ?
15:56:31	 From Nima Sedaghat : If one wants to summarize the method in a non-technical way, would this be right to say that: the idea is to partition the training set and only use those partitions that are appropriate given the test set?
15:58:59	 From Marc Chaumont : Ok !
15:59:01	 From Marc Chaumont : Thanks
15:59:01	 From sergey : Do you end up with multiple  models trained on each strata separately? 
15:59:51	 From Nima Sedaghat : Thanks
15:59:58	 From Dmitry Golubkov : how is the optimal number of strata chosen?
16:00:11	 From Tom Junk : What happens if the problem is a bias in the sampling of the labels in the training set and not a bias that depends only on features?
16:00:31	 From Samuel Centeno : But there is still a coefficient of error built into each stratum!
16:00:33	 From sergey : OK thanks!
16:01:19	 From Marc Chaumont : The weighs between the the strates predictors are fixed (learned) ? or do they depends on Redshift or another adaptation parameters (related to the input) ?
16:01:27	 From Dmitry Golubkov : thanks, so the number of strat is another hypoerparameter to be cross-validated?
16:05:12	 From Dmitry Golubkov : i see, thanks!
16:05:22	 From Marc Chaumont : What is the feature space (what are the features) ?
16:05:34	 From Miguel Cardenas : Have you evaluated the robustness of the approach by alterning (a little bit by the input data, some small noisy is enough)?
16:06:41	 From Marc Chaumont : Ok -> raw features + eventually hand craft features.
16:06:45	 From Marc Chaumont : Thanks.
16:07:39	 From Miguel Cardenas : Thanks
16:08:01	 From Shahbaz Ahmed Alvi : Thank you for the great talk and also the people for asking great questions. 
16:08:09	 From Miguel Cardenas : Great Talk
16:08:12	 From Marc Chaumont : Thanks
16:08:16	 From Samuel Centeno : Thanks
16:08:17	 From Frédéric Comby : thanks
16:08:20	 From Dmitry Golubkov : thanks