Player is loading...
Airworthy AI; challenges of certification, Q&A
Dr. Luuk van Dijk, Founder and CEO of Daedalean
Friday, Oct. 12, 2018 · 2:56 p.m. · 06m 33s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:00
comments i
00:00:04
yes the additive also use all the place
00:00:08
so
00:00:10
uh so you have a well established means of getting your p.
00:00:13
you're on your d. you know you have a model
00:00:17
um during your tuning uh and then you verify in flight but the
00:00:21
thing is stable your mobile and the holes ability to once you
00:00:24
have it and you have your mobile you can push stability really option of i think the whole the option of analysis was invented
00:00:31
is it but also impact is so there's that you so you have this controller you put it on the test then she
00:00:36
knew subjected to all kinds of uh of normal conditions you prove that it remains stable that's how these things work
00:00:49
even
00:00:51
so that's that that that goes along the line of would you say that why would
00:00:56
i replace a well understood good working p. i. d. controller with a neural network
00:01:03
to certify the network yeah but not to solve
00:01:08
yeah
00:01:15
long term process not sure i mean it it could be it could be good um
00:01:20
so the this paper from a nice and honeywell um on that the systems
00:01:27
talks about gain scheduling which is where you have so you're you're the controls really relies system
00:01:32
um that operates in the specific range you prunes billy there but then maybe go
00:01:38
to a completely different operating range you know you could put in different games
00:01:41
so this is what i figured oh you're gonna change your gaze n.
00:01:45
flights no now so then we're already at the boundary of um
00:01:51
what is a set that rocket science uh but not a good way i'm sol
00:02:02
it would be hard to outperform
00:02:04
no no actually would be very thing um
00:02:09
to have a instead of a bit controller have neural network outperform it completely that actually very good
00:02:18
thing
00:02:21
ha uh yeah let me think well that's more
00:02:33
okay anything about it some more
00:02:39
more questions suggestions and
00:02:47
i i i
00:03:00
that was exactly my question
00:03:04
here
00:03:06
so you have to use the show some form so our first hunch was i can give you some thoughts
00:03:11
about working here but i don't wanna be a sport i'll be a creative to you too much
00:03:16
so our first thought was you look at the entropy distributions of colours this abuse of
00:03:20
bigger features rather than if you for that have the relevant features are really
00:03:25
what gets triggered in a network so if we can compare what goes on in the networks and we can say given the network
00:03:31
these data sets trigger have the same kind of figure patterns on the inside
00:03:36
therefore for the purpose of the network the we are the same because they have the same sed statistic
00:03:41
not just uh the traitor but also the code for if you get all the pairs of nodes
00:03:45
and you make a that you get the distribution of a three ring together or not and they
00:03:49
are roughly the same for all players you can think of or a large sample of those
00:03:54
maybe these data sets are the same for does not work in any
00:03:59
meaningful way and that could be way to put back on that
00:04:10
oh
00:04:14
uh_huh
00:04:19
okay
00:04:22
well
00:04:26
well that would be very elegant
00:04:33
okay
00:04:45
yeah
00:04:50
yeah
00:04:53
ooh
00:04:55
yeah i would yeah i would hand craft this so that it looks like a accent 'cause one oh
00:05:00
it's actually very hard to get realistic lighting so we've been but so if it looks like plastic
00:05:07
i'm doing real time that we don't have to put some whatever that so and usually do
00:05:12
this would be to localise might have the to maximise local contrast and post rise and
00:05:33
oh
00:05:34
but we we could still that do that on top of that because we because we need to do
00:05:39
so i'm mega pixels are luxury um so we're going to use uh in in the lower levels
00:05:45
you have to um do some yeah get rid of the pixels that's the big one
00:05:53
i'm a reduced resolution of it so we have to do that anyway but then what
00:05:59
you say we can still do it the five twelve or five twelve later
00:06:07
and that that's what i propose i just to make the input space more regular lies them about it
00:06:21
or suggestions
Conference Program
Airworthy AI; challenges of certification, part one
Dr. Luuk van Dijk, Founder and CEO of Daedalean
Oct. 12, 2018 · 2:05 p.m.
350 views
Airworthy AI; challenges of certification, part two
Dr. Luuk van Dijk, Founder and CEO of Daedalean
Oct. 12, 2018 · 2:30 p.m.
107 views
Airworthy AI; challenges of certification, Q&A
Dr. Luuk van Dijk, Founder and CEO of Daedalean
Oct. 12, 2018 · 2:56 p.m.
181 views
Recommended talks
Concept Measures to Explain Deep Learning Predictions in Medical Imaging
Mara Graziani, HES-SO Valais-Wallis
May 3, 2019 · 10:32 a.m.
458 views