Open in new window / Try shogun cloud
--- Log opened Tue Jan 24 00:00:19 2012
-!- Netsplit *.net <-> *.split quits: @sonney2k09:19
-!- Netsplit over, joins: sonney2k09:22
-!- Ram108 [~amma@] has joined #shogun15:30
-!- Ram108 [~amma@] has quit [Ping timeout: 240 seconds]16:55
-!- blackburn [~qdrgsm@] has joined #shogun17:02
-!- Ram108 [~amma@] has joined #shogun17:22
-!- Ram108 [~amma@] has quit [Ping timeout: 248 seconds]18:03
-!- Ram108 [~amma@] has joined #shogun18:16
blackburnsonney2k: will you sign against elsevier? :)20:54
blackburnI would protest about every non-free shit if I was a scientist hah20:54
-!- mode/#shogun [+o sonney2k] by ChanServ21:44
-!- Ram108 [~amma@] has quit [Ping timeout: 244 seconds]22:53
@sonney2kblackburn, I thoroughly hate all this non-open access crap23:16
blackburnwow you are here23:17
@sonney2kNowadays I consider it good policy to not review nor publish for / in such journals23:17
@sonney2kblackburn, yeah but will go to bed soon23:17
blackburnsonney2k: I did not commited any idea yet sorry23:18
@sonney2kblackburn, did gsoc 2012 already get announced?23:18
blackburnno no23:18
blackburnnot yet23:18
@sonney2kI was working on the debian packages (the last days...)23:18
blackburnfor shogun?23:18
@sonney2kuploaded like 4 revisions...23:18
@sonney2kweird errors / warnings with clang...23:19
blackburnsonney2k: is there any way to automatize it??23:19
@sonney2kI don't understand why it fails on powerpc.23:19
@sonney2kblackburn, well we could have a buildbot building nightly debian packages for debian unstable23:19
blackburnwe could or we can?23:20
@sonney2kbut that will usually fail due to whatever stuff is broken in debian unstable :)23:20
@sonney2kit is a matter of typing one command on the cmdline but still work23:20
blackburnsonney2k: did I told you accuracy on these road signs?23:21
@sonney2kblackburn, at some point yes but no idea where you are now23:21
blackburnwell stucked on 96%23:21
@sonney2kdid you already use virtual images?23:21
blackburnno, not really23:22
blackburnI don't really want to use it23:22
@sonney2klike slight rotation / translation23:22
@sonney2kthat will improve another few percent23:22
blackburnI'm trying to find a way to localize images a little23:22
blackburnthere is a bad background23:22
@sonney2kblackburn, regarding GSoC23:22
blackburnmy descriptor is gradient based so I need to clean up images23:22
@sonney2kI think we should change the old website quite a bit23:23
blackburnyeah we def. should23:23
@sonney2kI mean I would want to have very clear focus23:23
@sonney2knot thousands of ideas like last time23:23
blackburnI would want to work on two things23:23
blackburn+ 10 more dim reduction algos23:24
blackburnor ECOC stuff23:24
blackburnwell I will have much time so I guess actually I will work on everything23:24
@sonney2kWhat I could imagine is: a) structured output learning b) multiclass learning c) input / output stuff d) cross-validation e) merging of SVM solvers f) opencv integration ...23:25
@sonney2kit sounded like chris wants to mentor domain adaptation algorithms, we might have a mentor for gaussian processes too23:28
blackburnwhat the f is domain adaptation?23:29
@sonney2klearn on data A but apply the method to dataset B (which has a slighly different statistics - so adjust for that)23:30
blackburnwell I would be interested too23:30
@sonney2kwe have stuff like this already in there23:31
@sonney2kblackburn, this year is much easier for shogun23:34
@sonney2kwe have buildbots and a better structure23:34
blackburnyes, much better23:34
@sonney2kso all these things are just more or less minor extensions23:34
@sonney2kthat all integrate rather well23:34
blackburnand at least I will help you much ;)23:35
@sonney2kI guess our chances to get in are not so bad given that chris di bona liked my summary :)23:35
blackburnhah yes, we had some fame23:36
@sonney2kI wish we would find someone to do a) structured output learning b) multiclass stuff (I am particularly interested in this - the nips ecoc/tree paper sounded fast + accurate)23:37
blackburnmy gf bachelors thesis would be about ecoc hah23:37
blackburnwork/thesis/no idea how to call it properly23:38
@sonney2kI also would wish cross validation could get some syntactic sugar and more wide use23:38
@sonney2kI hope Vojtech will mentor again this year23:39
blackburnwhat would he mentor?23:39
@sonney2khe could really help to get all these SVM solvers in23:39
@sonney2khe is the ueberexpert there23:39
blackburnin SVM solvers?23:40
@sonney2kyes - he is the author of Ocas / GMNPSVM, ...23:40
blackburnsonney2k: I still feel unhappiness mixed with depression when looking into svm folder23:40
@sonney2kthat won't change. way to difficult/time consuming  to rewrite these things23:41
blackburnI hope I would see any progress in my life hahah23:41
CIA-10shogun: Soeren Sonnenburg master * r210ae0a / (4 files in 3 dirs): fix various errors/warnings occurring with clang -
@sonney2kahh btw, some nice tutorial would be a *great* task too23:41
blackburnI hope 40 more years would be enough23:41
@sonney2kso applications etc23:42
blackburnI believe student can't write nice tutorial..23:42
@sonney2kwhy not?23:42
blackburnif it wasn't phd student = expert in machine learning23:42
@sonney2ktask could be to define some nice data set based on mentor input23:42
@sonney2kthen run all alogirhtms23:43
@sonney2kalgorithms on that23:43
@sonney2kwith nice images23:43
blackburnwell only if he was really educated in ml23:43
@sonney2kand write text + formulas that he can find in the code / papers23:43
@sonney2klast year we had such ML experts23:43
@sonney2kwe didn't accept them23:44
@sonney2kPhD students...23:44
@sonney2knot sure how many students we can mentor this year23:45
@sonney2kprobably around 5 again - not more23:45
@sonney2kless is better if we find really good ones23:45
@sonney2kof your type btw :)23:45
blackburnhah thanks23:46
* blackburn is playing with opencv23:46
@sonney2kbtw parallelization is also sth23:47
@sonney2kbut I guess should be done by me23:47
@sonney2knot a student23:47
blackburnsonney2k: no, parallelization should not be done by student23:47
blackburnhe would stuck and etc23:47
@sonney2kyeah, btw boosting algos are still missing from shogun23:47
@sonney2kI would love to have some of these too23:47
blackburnand I consider doing vectorization not parallelization..23:48
blackburnbut I still have no machine with good graphics23:48
blackburndon't really want to buy new one for that :D23:48
@sonney2kI am not a big fan of this GPU stuff yet23:50
@sonney2kjust have no applications where GPUs really could help23:50
blackburnsonney2k: well my stuff would run much faster with gpus..23:50
@sonney2kok I think we have a much better picture now what we could do23:51
@sonney2kso it is easy to write something up23:51
@sonney2kblackburn, yeah you have some quadratic/cubic algorithms in there23:51
@sonney2kthat certainly helps23:51
@sonney2kand all nicely gpu-parallelizable23:51
blackburnsonney2k: let multiclass stuff be done by me in feb-apr23:51
blackburnI would need it anyway23:52
@sonney2kmulticlass stuff involvement can start with simple stuff but can be arbitrarily tough23:52
@sonney2kso a student could as well do the harder stuff23:52
blackburnsonney2k: well I really need it so I'll try to get it done even in feb23:53
blackburnincluding NIPS things23:53
blackburnI'm pretty scaried with it but I hope to get through hah23:54
blackburnsonney2k: for me would be nice to work with chris, we could write one more article :D23:54
@sonney2klooking forward to this code - I will probably be your first user23:54
@sonney2kthere are some ecoc libraries around already though23:55
blackburnyes I am aware of it23:55
blackburnsonney2k: I was thinking about generic ensembles and generic multiclass machines23:56
@sonney2kthat will involve quite some refactoring23:56
blackburnit should be done really *generic*23:56
@sonney2kwe have general OVO / OVR in MulticlassSVM23:56
@sonney2kat least for classify23:56
@sonney2kso one could probably get rid of LibSVMMultiClass or so23:57
blackburnno, I mean something like constructing trees manually23:57
@sonney2kfor training (which just does OVO)23:57
blackburnecocs, etc23:57
blackburnand not only svm23:57
@sonney2kblackburn, the code is not just for SVM23:57
@sonney2kit is only encapsulated in MulticlassSVM23:58
blackburnsonney2k: one more idea!23:58
blackburna lot of23:58
@sonney2kjust because we had no multiclassmachine23:58
@sonney2klinear regression is totally missing23:58
@sonney2kand these are quite some low hanging fruits23:58
blackburnbut I believe we should be 'minimized'23:59
blackburnwe should not implement everything..23:59
blackburnmahalanobis distance could be gsoc-entering task hah23:59
--- Log closed Wed Jan 25 00:00:03 2012