Transcriptions
Note: this content has been automatically generated.
00:00:02
a professor should grow we a slab and with my quarters work on to accept
00:00:07
um so today if it kind of safe if you go to
00:00:10
an airport and pass although security checks done by humans
00:00:15
should you feel safe windows security chit checks we did on
00:00:18
the machine learning algorithms artificial intelligence well yes and no
00:00:22
i know if i tell you that today uh terrorists country d. printable changed
00:00:27
if you pick sets a few meters in the colour of the bob
00:00:30
and make a classifier in the airport say this is a cute find this actually happened with a classifier
00:00:36
and yes a a a because more and more people and researchers are working on
00:00:40
the field i sit in in particular here at e. p. f. l.
00:00:43
we've uh come up with a couple of solutions to the vulnerabilities of machinery machine
00:00:47
and it has a lot of unabridged i'll talk about one in particular average
00:00:52
machine learning relies heavily on average and if you have done some basics astrology you'd know that
00:00:59
averaging is the worst way to compare two populations for example if you take the g. d. p. of denmark
00:01:04
averaged over the population of that market gets a lower value if you do the same averaging over the
00:01:09
us population and no one is full to the to say that the typical us citizen is richer
00:01:14
or comparable to the done and then use it is the same for this room is the is the billionaire in the room
00:01:19
the average income of this room is a hundred million that's this is a very bad way to classify rules were academics
00:01:27
so uh of course machinery resources no an alternative to average of the media
00:01:32
the meeting is very easy in single dimension of uh of variables you just take the
00:01:36
bible's rank them and take the one that separates the population into two parts
00:01:40
and this would be robust classifier that's determined that machine learning gives in very
00:01:44
high dimensional space is a models are vectors of a hundred billion parameters
00:01:50
and you can just rank them and take the how the this value that separates the need to to hop operations
00:01:55
so we michael cocoa cultures would come up with a with alternatives
00:01:59
that are practical and that we proved mathematically are safe
00:02:03
uh both alternatives i'm somehow inspired by the median but others
00:02:07
are lip shoots filters or uh other utterances we've
00:02:11
presented our results uh to the machine and you'd you so uh they are kind of
00:02:15
accepted and now uh we are implementing that as a system on top of concerts
00:02:20
i would also tackled some other topics in i. c. t. like safe interrupt ability and three
00:02:24
years this can happen inside a neural network at an individual neural more on uh
00:02:30
a level and if you're interested in details and the details of this works so