Player is loading...
Q&A
Prof. Masashi Sugiyama
Tuesday, May 28, 2019 · 11:59 a.m. · 06m 50s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
yeah it's less oh yeah yeah yeah
00:00:05
i see i think ah it's really like the same brought in to go back 'cause
00:00:13
going this way it's been years so it's nice to be away from that direct on
00:00:19
uh_huh uh_huh uh_huh
00:00:25
right is the biggest that maybe we came from different direct sun but maybe that that's
00:00:29
one possible various and i i'm not sure that it really makes a lot of difference
00:00:34
but in um i'm playing with this and we just decided to step back a 'cause we didn't have any memory for for what i read in
00:00:40
yeah but but we can memorise and like reading maybe it's no problem
00:00:44
i don't know how how big a difference it is but i think the essence idea it's quite the same
00:00:51
yeah thank you and yes right yeah oh no no
00:01:01
in any case yeah they have you have ideas some some yeah uh_huh
00:01:07
yeah ah so uh like pixie i i worked on
00:01:11
the problem called come are it's if an actor also published
00:01:14
a book on that it's a kind of trance for learning like training domain and this domain have difference you do something
00:01:20
and we can somehow combine those techniques she are 'cause as i say this is the simplest impeccably skin my this
00:01:26
on like we can also like of some importance weightlessness and apply for like coming up this and this is also possible
00:01:33
and also in some applied project we have already done and it is martin in some cases and it is not in some other cases
00:01:40
but uh this is the simplest expression and becomes a
00:01:42
combine any like fancy techniques developing wasn't running in this film
00:01:50
it's ah it's
00:01:59
you know oh uh_huh yeah ah i guess
00:02:12
now i i completely agree with it so we we should have done something for that but so far like for that can communism
00:02:19
we just did some applied without and we just focus on the simplest for mark but right so maybe we should also the that
00:02:27
yeah thank remote
00:02:29
yes
00:02:33
it's it's
00:02:39
yeah uh_huh
00:02:48
oh you are so uh this is enough week label week data exactly quite ambiguous also you might though so i i
00:02:55
maybe my does not really right but like okay i not
00:03:00
fully need a bit data but we suppose something is missing
00:03:04
like in this in this case negative labels i'm missing the dusty you learning maybe p. n. is bit
00:03:10
different but then pecans case we we uh missing negative
00:03:14
data but posed to be tested before they could it
00:03:17
posted confidence so maybe the uh not technology is not that really consistent here but we are trying to say that
00:03:24
thankfully supper was linings extremely expensive in many cases but
00:03:27
in some applications like collecting posting data it's quite cheap well
00:03:31
correcting typos that convinces almost for free then we should be going to use that information to training class right yeah
00:03:38
so we're not really saying like these methods the university useful for almost all problems
00:03:44
for some problems like i it's maybe useful like for example okay we're working on
00:03:49
brings ignore analysis for some purposes in that case i. t. was do they thought
00:03:55
is this a common for free so we aren't sure about like data collected in this race all always posted
00:04:02
then posed to the car can be clicked it for free and and the the data is also for free
00:04:07
so then basically in terms of the cost it the same ostensible was
00:04:10
lined but we can be performed syllable was done with a with a guarantee
00:04:15
the that the kind of goal we are heading for us to it maybe we we're not we summarise things quite well but
00:04:22
uh i i mean i don't read yeah i answer your question created but
00:04:31
oh
00:04:35
yeah
00:04:38
yeah yeah yeah yeah yeah yeah that's with the p. confiscation i think yeah yeah that that he
00:04:46
can for data and also if you have incomes they thought that that can also be just you know
00:04:52
okay uh_huh
00:05:00
yeah yeah i i i ah mm
00:05:17
uh_huh
00:05:22
ah okay uh_huh yeah i agree with that that that's a very important point and we had a three so some
00:05:32
new play pause on only the bias this jason like we have biased 'cause they they tell a biased native data
00:05:39
uh oh 'cause we need some additional some some day uh but we can also some somewhat cope with that
00:05:43
kind of station yeah we also only of that kind of seasoning pack this yeah but that's really important i think
00:05:50
yes yeah i
00:05:57
a what
00:06:05
yeah so okay uh you are really right so we had the bomb the like this
00:06:10
but he actually i just showed all that but there's a hidden constant it came minus one case the number of classes
00:06:17
the because it's clear like if it's k. classes like one label that one ordinary label
00:06:22
request one thirty minutes long company the labels and really had a pound in that way
00:06:27
but yeah i this is not not found actually tried to improve it was maybe if a mini class it's some are correlated
00:06:34
if we can take that correlation into account hopefully we can be just a
00:06:37
minus one pop up to a small number but maybe you didn't write up here
Conference Program
Recent advances in weakly-supervised learning and reliable learning
Prof. Masashi Sugiyama
May 28, 2019 · 11:04 a.m.
655 views
Recommended talks
Large scale image annotation: learning to rank with joint word-image embeddings
Samy Bengio, Google, USA
Sept. 1, 2011 · 1:42 p.m.
110 views