Player is loading...

Embed

Embed code

Transcriptions

Note: this content has been automatically generated.
00:00:01
prove that the neural network in a certain application to save
00:00:07
so before i spoil the creativity with my hunches
00:00:12
maybe somebody has worked on this plan to work on this
00:00:19
simulated doable stuff
00:00:31
okay so you take it toward version of the problem you have the neural nor networks
00:00:35
all that and show that it's all that as well as the artificial handcrafted one
00:00:44
right um so we did have this simulation simulated input as ground truth as it
00:00:50
is a major ingredient in in any approach we take a any other
00:00:56
well well discussed this at length does the rest of my talk them but i didn't mean to talk happen are
00:01:03
about this but so far um so let's first narrower ourselves
00:01:06
down to this reference system um it's it's uh
00:01:11
uh actually something we are building um so we have cameras
00:01:14
coming in we have a pretty high resolution picture
00:01:19
oh coming in we pretty process it so that goes to lower resolution we deal with the artifacts in the image
00:01:24
like bad pixels and stuff also we post rise the bit so that the neural network already sees
00:01:30
something that looks as if it was in again computer rather than the full spectrum of reality
00:01:35
uh which we hope will help later then go through the uh resonant thirty
00:01:39
four or whatever we once and it outputs the segmentation of the image
00:01:42
um we have two major applications one is recognising landing strips
00:01:47
um so there's the the black things what what lines in the middle of the circles with age they're relatively easy uh an
00:01:53
extension of that is anywhere look anywhere and see the difference
00:01:57
between saving and save 'em but this is the more
00:02:02
tangible it's more an edge case 'cause we can actually see if you can see lana strip or not um
00:02:08
uh and the other one is uh recognising things guy like you drones and birds
00:02:12
non cooperative traffic we call them um it's important because uh humans actually
00:02:18
i have not been of i. sites to sport a small drone at six hundred
00:02:22
meters or six seconds away if you're travelling hermes second um so the same
00:02:28
the same infrastructure uh same hardware very somersault person
00:02:33
network topology that in one case the false
00:02:35
positives or obviously far worse than the false negatives the other one we'll rerun yes
00:02:44
ah very good question that's why we have um well
00:02:49
first of all good they don't have to they just have to pass the exam
00:02:51
fume pilots we have um a corporation with the that and it's really
00:02:57
uh a whole chiffon competition shove them into two or they
00:03:00
have a department dedicated to avionics and they're exactly
00:03:04
um uh i do a project like this to figuring out how good humans are this at all
00:03:11
um so you could say well as soon as we prove that we're better than this set of humans over there in the lab
00:03:17
we we had done but we'd have to go to your thirties and they say that that's very nice
00:03:22
lots large story and still you have to um be better that you would have to
00:03:27
be as good as we expect from evaluation systems and not as humans um
00:03:48
yes so these companies that have been done so humans are actually dollar beam and best
00:03:55
uh and so a a a human uh messes up or once per utterance is seven hours
00:04:01
ah to uh if you have to humans is not the square because they can talk
00:04:06
to each other so and all these design flaw and uh and the cockpit
00:04:11
uh is that a pilot can talk is are actually it airbus internal joke is to replace the
00:04:15
second pilot with a dog will bark at the first part of the price thirty but
00:04:21
ah i'm there could be a credible step on the way to full autonomy um
00:04:26
so um so if you just look at the isn't database
00:04:32
you see uh uh the probability of a pilot dying
00:04:35
uh wise on hand is this number so we know and then afterwards there's the analysis of what went wrong
00:04:42
i think the fundamental problem is there that's not a very huge database because there's not not that many planes flying is not many
00:04:47
people trying not many hours of flight have been not ten to
00:04:51
nine hours a lot of hours um it's clearly a
00:04:57
yeah um if you have a very very experienced pilot with twenty thousand
00:05:02
hours of experience twenty thousand flight hours that's very experienced pilot
00:05:07
um this couple things we can do better though uh we can share our knowledge uh which
00:05:12
humans can only do you know over beer or in an informal setting like this
00:05:17
um we can pull the learning um that we can do off line learning is also something we
00:05:23
can clearly do better than humans we can record everything for later do learning online and um
00:05:30
uh well i'm gonna have myself so this is our references and i won't
00:05:33
talk about this uh as of now is the system for which we
00:05:38
like to go to us and say look we improve the thing that was good enough therefore the system as a whole is good enough
00:05:43
so um preprocessing so it's it's sort of a game world in
00:05:48
reality remembered segmentation then this is the single shot network
00:05:52
if you're so you have any recurrence or any buildup of stake in there
00:05:56
it becomes a are dependent on the history of images you show it
00:06:00
i think i have no idea how to track any how to make any
00:06:04
meaningful statement about adding more so let's focus on easy one picture
00:06:08
uh one segmentation combat but we can you post processing so with every
00:06:12
new frame of our fifty frames per second is a complete
00:06:16
surprised to see in the the to the neural network and it comes up with an independent estimate
00:06:21
after you moved for a couple of seconds you have completely independent image anyway
00:06:26
and we can do classical radar tracking algorithm new common field of the test
00:06:30
or something there i expect some consistency of uh of the object or
00:06:36
if you apply this to wire detection wires on major cause of accidents in uh in flying
00:06:42
i'm a sinner just seen the wire ones right we can have a post processing that remembers that
00:06:49
it's all wire there and until we are absolutely sure was mistake we will not go there
00:06:55
that's the way that you can increase the cost of the the the president nicole although the whole
00:07:01
system um even if you're no we're gonna middle has a finite chance of getting it wrong
00:07:07
um so the important constraints that we set ourselves now we we restrict ourselves to the corner
00:07:13
of the design space where we have no adaptive online learning so we will not have
00:07:17
uh we we're not trying to make an actual biological entity that
00:07:21
wants to fly 'cause turns out the make terrible pilots
00:07:24
um uh we are not in control look with skirts or the simplifies the notes of the whole thing and
00:07:30
we have no buildup of states we have no recurrence um in the network that makes it remember anything
00:07:37
yeah if we can do this we're already we can do pretty cool
00:07:41
things um okay so if you go over the way that a
00:07:46
f. a. any other reason about this gives me but uncertainty any
00:07:50
uncertainties caused by randomness um i mean this is not an
00:07:55
deep statements so what are the sources of randomness in uncertainty in our neural net and the application thereof
00:08:02
so the most obvious ones that can be noisy input and we've all heard the is fantastic
00:08:07
examples of um that was your attacks we try to find a little noise pattern that
00:08:11
makes the outcome a completely different um so we kind of any of that
00:08:17
um that's the only thing that really happens at run time for the system and uh and inference time
00:08:22
um the rest are all things related to uh the training
00:08:27
or the design phase as a airspace would call it
00:08:31
so far as we see it the first problem is that your test set
00:08:36
is the sample drawn from reality and you have the sample problem that has to be representative
00:08:42
the fundamental problem for anything you testify nuts at first
00:08:46
actually worse because they're on there isn't that much
00:08:50
footage of planes flying in extremely dangerous conditions not only did you not tend to carry recordings the
00:08:58
recording the there's no video and see no it doesn't happen that often it's very rare events
00:09:03
statistics um these accidents um then when we get to the second step so this this
00:09:10
meal data is far too valuable to waste on training the network you need
00:09:14
or the more magnitude than we have available for testing at all so you wanna
00:09:18
train is exclusively in simulator we think that's the only feasible way so these
00:09:25
training sets are synthesised for randomness actually the better the randomness
00:09:29
the better they are but how do you so there
00:09:33
you would have to argue that the randomness leads to more certainty um which is something we have to get a grip on
00:09:39
and then you run the my eyes the training processing different runs
00:09:44
we huge you would end up with different neural networks
00:09:47
um so are they meet different uh in which case that could be totally
00:09:51
valid and they could be multiple so good solutions to the same problem
00:09:55
or maybe they're essentially the same because there's a high degree of symmetry
00:09:59
representational network usually graph um user what's it um um so
00:10:08
if you
00:10:10
even adam excesses couple times even if the end result is a good no system to argue that is good enough you have
00:10:16
to open up the box and show you understand what's in there so that's the channels we have to sort ourselves
00:10:22
known problems for the design we're talking about his determinism sir
00:10:26
showed the same image i get same output will probable
00:10:29
um we excluded ourselves from modifying the weights in flight because
00:10:35
no that would be a very bad idea for all kinds of reasons not only would be hard to prove it safe enough i also think it would be safe enough
00:10:42
so it's good that we can't prove it safe enough because we'd be lying um
00:10:46
buildup of states is also something we avoid and go to evaluate the network
00:10:53
that's just you know well written c. plus plus we not to do that actually we
00:10:57
did it because the the first computer i showed you comes with a very ancient
00:11:01
um cheap you for on which tend to vote doesn't work so we roll around here where the implementation of that there's a
00:11:07
flow uh evaluation library which is you know classical solid hardcore
00:11:14
writing of software that we know what to do so
00:11:19
um the run time determinism that was the first court problem actually i've
00:11:24
oh i'm gonna go over these four points i mentioned earlier briefly um so the determinism is nice but we need
00:11:31
stability really if we change the input a little bit we should not get too widely different output
00:11:36
now my hunch is no i'm just all theoretical physicists that
00:11:42
stumbled into this field but if the whole thing is really different trouble seems to me you have everything you need
00:11:48
mathematically to come up with the concept of d. outputs divided
00:11:52
by the input and put some buttons on this
00:11:56
anybody has a theoretical framework to do work on this does not um you read a lot about
00:12:04
uh at the serial methods that try to um a full network can use these to exclude
00:12:11
such things to uh is there a way to uh learn and the
00:12:14
sense star is um um yourself to noise um uh or
00:12:21
could ensemble methods being used here to get greater stability and also
00:12:26
to get a grip on what a stability means either questions
00:12:31
and somebody ah
00:12:35
it's
00:12:39
oh
00:12:44
oh
00:12:55
no so my question one what are the mathematical right
00:13:02
and so the old really you see get this training set you have this
00:13:07
test set a you measured the gene recall and they are this number
00:13:13
right there this number on the test set is this test sets efficient representation of reality um
00:13:22
oh well maybe i know how would we go about proving that
00:13:37
i i i
00:13:47
yeah
00:13:58
so
00:14:01
yeah
00:14:17
exactly that is a pro as i said i've come here to ask questions not to get
00:14:23
yeah um
00:14:36
oh
00:15:02
uh_huh
00:15:21
i
00:15:41
i
00:15:49
ooh
00:15:54
right
00:16:02
i think this is a call to arms for for no research groups that are looking for the next cool thanks
00:16:09
um basically so basically what we want avoids is that somebody can
00:16:14
put an orange or pink ball in the corner of the
00:16:17
field of view of the aircraft to make you think that there's a landing site over there that would be that
00:16:26
okay well uh that would be nice
00:16:39
so we could start and also so if we have a bunch of these uh uh networks trained separately when they are separate
00:16:46
uh we can maybe use at all get to that later so this is the only one where it's a property of the networking
00:16:52
comes out that we apply in there that we wanna have properties of that applies on time the rest is really about
00:16:58
opening the uh on the box of the the design phase and
00:17:04
and uh uh that those were the the the sampling error in the test set i think the air
00:17:12
uh we might actually have a chance so we're not gonna get out into the nine hours of flight
00:17:18
uh but for the case of the landing strips there's probably fifty thousand landing strips in the world
00:17:25
so we could probably get photos of forty thousand of them and then we can par
00:17:30
parameter tries what they look like by you know how far away which angle
00:17:34
uh we're looking uh and uh and then we can take a very large sample those or um
00:17:41
uh and then we have to multiply this by what they look like in the weather soul
00:17:47
there's by relevant parameters of the weather is to the
00:17:50
uh i'm really relevance uh aspects of the
00:17:55
lighting conditions it's the translucent c. of the air and how much letters to begin with
00:18:00
um so that's seven dimensions is space that you can quantify and
00:18:04
we can take samples and and then we can see
00:18:07
if we have a over or under sampled hard parts so
00:18:12
i have good hope there and so it's not even
00:18:14
that big data um fifty thousands times say um
00:18:20
thousands a weather conditions that is uh sounds like a big number but currently there's
00:18:26
a five hundred hours of video get uploaded to you'd you were minutes
00:18:31
uh they're not all about lying but um that means that and the seven hours actually
00:18:39
i've got forty days you'd you workloads um is that thirty the the factor
00:18:44
is real time to upload it is thirty three thousand or something so
00:18:51
um so we don't even call a bigger setting or a means that to go get it but um and then maybe we
00:18:58
can make some statements about the quality and the the variance in
00:19:03
the um sample set based on bootstrapping and i know
00:19:09
i said patients is that stations in the audience that say all we always do like it
00:19:15
okay actually uh um the more existing one is that uh as i said impact is known appeal data to train on
00:19:23
and training uh to put it differently it would be a giant waste to uh use real data
00:19:30
for training um we only one isn't protesting so we wanna set aside all the training
00:19:35
and then it's nice if any and it works all the testing set but a asses gonna ask so if
00:19:41
you traded on that's you know how does that if you if he's synthesised into the nine hours there
00:19:48
how have you not that on something could be relevance and
00:19:51
um why uh should we just that this means anything
00:19:58
so what have you thought this network um and specifically how have we
00:20:04
made sure we don't over fit on synthesise aspects of the
00:20:09
uh of the of the training set that we happen to not speak up on
00:20:14
in a um in the real data uh does that uh um again
00:20:21
so these uh the serial things seem to be older age um can we uh the
00:20:26
somebody suggested uh you approach where you learned a on a
00:20:31
i think you told me this uh you have this uh the serial
00:20:34
had on uh on your network we can tell the difference between
00:20:39
uh i'm real and fake and you only learn by rivers propagating
00:20:44
um the printer network the features uh that we can use to distinguish those
00:20:50
uh we already have a very obvious that a preprocessor makes
00:20:55
the will look a bit more like a game engine
00:20:58
uh in real time so the the real time application the network see something that looks very much like a game engine
00:21:05
i'm more for the mentally ill we have to define some for similarity of the training and
00:21:11
the test set if we wanna say you know it makes sense that this test set
00:21:15
or objectively tests will be trained uh because they are the same in this way so what
00:21:20
does it mean for the sets to be different or the same that is um
00:21:25
an interesting concept to have a number on um then
00:21:32
so if we if you have a p. i. d. controller and you have a process the set the uh the three
00:21:37
gains and every time you run it comes up with three different numbers at all sort of work the same
00:21:43
yeah it's like a big enough to convince someone to just there a jet engine control to it so you
00:21:49
wanna show that if you run the that work several
00:21:52
times and you do get different results they are
00:21:56
different for good reason reason under their objectively still good solutions
00:22:00
uh or that there are essentially the same modulo um permutation of all the weights and
00:22:06
all the uh um uh act activation all the connections so um uh um
00:22:16
if we do have
00:22:18
objectivity good but this several our networks uh which you could maybe force by
00:22:24
severing the making different apologies uh if we can some approve the
00:22:28
oars independent that actually you strong ass it be tool 'cause we could
00:22:32
put seventeen of them in a box in the committee no sample
00:22:35
and we have the expected amount of disagreements and essays only all agree
00:22:40
uh then there's probably common mode are that's the typical fold in these if you only systems is you have these
00:22:45
we don't thence with this we don't see built in but uh there's still way that they can all
00:22:51
phil at the same time because the uses one thing that's the same for all the all the common mode are
00:22:57
and people are could be allergic to it so if you have a fault in a in aviation
00:23:02
you know the rude call you in everything is brown and so you have the
00:23:05
root cause analysis isn't root cause analysis that some common common mode failure
00:23:09
people are gonna be upset um if we have an ensemble and all this and they all agree and of sense
00:23:15
it was probably something fishy going on at the same time if they all disagree so all of the seventeen
00:23:21
you now have the predictions all over the place that
00:23:24
that's also a good intrinsic a measure of um
00:23:29
uh how we should trust is doing on time so we can use the mortar during
00:23:33
run time how much we just the system um then um so to summarise
00:23:41
if we wanna convince the f. a. a. that it's good enough it's not a
00:23:46
good enough to show that it works on this test that because the test
00:23:49
of this small song good enough to show that we can do better than humans humans aren't that good and on so many of them and it's not
00:23:55
that many hours of flight we really have to open this box and see what goes on the inside so that we
00:24:01
can compare networks layer by layer and say this is means same as this and this is what it means
00:24:07
uh and so that we can compare data sets um given the same network perhaps uh to say
00:24:13
but we've learned here the e. is a generalisation of uh what we measure there
00:24:20
um then random thing that i want to put on some slides if you start relied retrain network
00:24:25
uh it will have been it was seen lots of things
00:24:29
that you never actually see in your training or your
00:24:32
television set and in fact this is will be get nodes that you could optimise out um to get them
00:24:39
nice once more network what you would do in aviation application i would propose you actually
00:24:45
you don't who come up to the next layer you will come up to um
00:24:49
a separate monitoring equipment and as if these nodes only fire you have seen
00:24:54
something that you haven't trained on so are a big thing anyway
00:25:01
um to wrap it up we needs to get a grip on reprise ability of
00:25:05
the training process if we wanna build the argument that we have trains
00:25:11
these fittingly numbers uh adequately so we need to have a framework
00:25:16
was that what goes on inside network so that we
00:25:18
can compare these networks and we credit sets and as far as i see it that's all we have to tackle
00:25:26
so you're helps a lot

Share this talk: 


Conference program

Airworthy AI; challenges of certification, part one
Dr. Luuk van Dijk, Founder and CEO of Daedalean
12 Oct. 2018 · 2:05 p.m.
Airworthy AI; challenges of certification, part two
Dr. Luuk van Dijk, Founder and CEO of Daedalean
12 Oct. 2018 · 2:30 p.m.
Airworthy AI; challenges of certification, Q&A
Dr. Luuk van Dijk, Founder and CEO of Daedalean
12 Oct. 2018 · 2:56 p.m.

Recommended talks

Décision assistée par ordinateur: de la donnée à l’information
Dr. Dominik Aronsky , Université Vanderbilt (USA)
7 June 2013 · 11:58 a.m.