00:00:06[Music]
00:00:09cool I’ll get started I’m Diego I work
00:00:14at Google brain but that’s all I can
00:00:16tell you my normal rule is if I told you
00:00:19I’d have to kill you but there’s a whole
00:00:20lot of you guys which makes that kind of
00:00:22impractical right now a standard
00:00:24disclaimer everything I say reflects my
00:00:26own opinions it’s not representative of
00:00:28my employer I have not been at Google
00:00:30long enough to know any secrets though I
00:00:32will admit that I’ve pilfered some
00:00:34publicly available slides so if there is
00:00:36a Google bias it’s not because they made
00:00:38me do it it’s because I’m lazy as far as
00:00:42some background for myself
00:00:44I broke a 13-year losing streak for the
00:00:47Philippines in the International math
00:00:48Olympiad I got the top prize in the
00:00:50world in the inter display compass in
00:00:52modeling and if you’re familiar with
00:00:53kayo competitions I also want one of
00:00:55those so I like to compete and do all of
00:00:59these kinds of things and hopefully that
00:01:01convinces you that I know that I’m
00:01:02talking about but let’s get started the
00:01:07presentation is deep learning what is it
00:01:09and what can it do for you
00:01:11but I think the very first question is
00:01:13why should I care and that reminds me of
00:01:16a story a machine learning researcher a
00:01:19cryptocurrency expert and an Erlang
00:01:22programmer walk into a bar
00:01:23Facebook buys the bar for twenty seven
00:01:25billion dollars and also another
00:01:29disclaimer you may not know this but I’m
00:01:30both a machine learning researcher and
00:01:32from San Francisco that means all of my
00:01:34information comes from Twitter that’s
00:01:36not a joke so prepare for that for my
00:01:39slides but back to why you should care
00:01:41machine learning artificial intelligence
00:01:43deep learning they’re all getting a lot
00:01:44of press these days they’re all doing
00:01:46lots of stuff and there’s lots of hype
00:01:48lots of news articles about all sorts of
00:01:50things and everyone seems to want to get
00:01:53into it but people don’t really know
00:01:54what they’re talking about it seems the
00:01:56only thing everyone’s really sure of is
00:01:57that artificial intelligence and it
00:02:00seems like very recently in particular
00:02:02deep learning will be a catalyst for a
00:02:04lot of change that’s happening and
00:02:06people are asking any questions all the
00:02:08time about this kind of thing and some
00:02:10of these recurring themes are how will
00:02:11the world change
00:02:12what can a AI and in particular deep
00:02:15learning do and how do I take advantage
00:02:17of these trends
00:02:19even the legendary programmer Jeff Dean
00:02:22has said if you’re not considering how
00:02:24to use deep neural nets to solve your
00:02:25problems you almost certainly should be
00:02:28it almost sounds like a threat either
00:02:31way I hope you’re motivated to learn
00:02:33that’s all I have for motivation so
00:02:35let’s get started into what is deep
00:02:37learning and it’s quite easy this is
00:02:41deep learning you could memorize this
00:02:44diagram this will make the presentation
00:02:46a lot easier so that’s basically it just
00:02:52kidding this is neither complete there’s
00:02:55a lot more to it than that and it’s also
00:02:57pretty complicated we’re gonna start
00:02:58with something much more simple namely
00:03:01calculus may not sound right it’s even
00:03:05chapter 1 of the book calculus made easy
00:03:07it’s titled to deliver you from the
00:03:10preliminary terrors but we only need a
00:03:12little bit of calculus and particularly
00:03:14we need an algorithm called gradient
00:03:17descent the derivative of one thing with
00:03:20calculus tells you is how to take
00:03:21derivatives and a derivative loosely
00:03:24speaking if it tells you how a functions
00:03:27output changes when you change its input
00:03:29and gradient descent is just moving
00:03:32along the direction of the derivative in
00:03:34order to minimize a function that you
00:03:36can take the derivative of so the
00:03:38insight into all of machinery Sigma Xin
00:03:41learning is figure out how to frame your
00:03:44problem in such a way that what you care
00:03:46about is differentiable or a proxy of
00:03:49what you care about is differentiable
00:03:50and then minimize it and this like one
00:03:55extremely simple equation summarizes
00:03:57almost all of the recent work in machine
00:03:59learning that’s happened in the last
00:04:03half decade roughly there’s obviously
00:04:05exceptions to this rule but majority of
00:04:08what’s done either has been done with
00:04:11this exceptionally simple thing or can
00:04:13be done with this exceptionally simple
00:04:14simple thing and that’s basically it as
00:04:18far as what deep learning really is in
00:04:20its essence
00:04:21this might sound well good but this is
00:04:24just machine learning when’s it become
00:04:25deep learning also easy it’s when you
00:04:28make it really deep it might sound like
00:04:31a joke but acts
00:04:32really what happens when you have these
00:04:35you know these machine learning used to
00:04:37be composed these very very simple
00:04:39functions because this is all we knew
00:04:41how to optimize and what happens when
00:04:43you stack multiple of these simple
00:04:45functions together you get something
00:04:46that’s much much more powerful if we
00:04:49don’t really know how much more powerful
00:04:50it is some might claim it’s
00:04:52exponentially more powerful but either
00:04:55way we know it’s much more powerful and
00:04:57simply stacking these things and using
00:04:59the simple algorithm is what’s caused
00:05:01the deep learning revolution to hit and
00:05:03it’s just using same old simple
00:05:05algorithm even though as a caveat to
00:05:09that we part of the hard part of deep
00:05:12learning is knowing that the simple
00:05:13algorithm will work for these very
00:05:15complicated models that have like stacks
00:05:18of layers and making problems non convex
00:05:20is that it for deep learning yes this is
00:05:25it roughly I’m skipping all sorts of
00:05:29details that I’m sure will be covered
00:05:31later from software that makes things
00:05:33easier to write like tensorflow or pi
00:05:35torch to hardware that makes things
00:05:37faster to run like GPUs or multi-core
00:05:41CPUs or TP use to commonly use some
00:05:44functions or layers that people have
00:05:47done lots of trial and error and just
00:05:50found to work well for some of today’s
00:05:52problems without very good justification
00:05:54and also commonly use combinations of
00:05:57these layers or architectures that
00:06:00similarly people have used trial and
00:06:02error and found to work without very
00:06:04good justification but for the purpose
00:06:06of that talk the simpler algorithm is it
00:06:10next big question is what can it do for
00:06:13you
00:06:15this may there was a recent paper that
00:06:17came out that makes it a little bit
00:06:18easier to answer because this article
00:06:20came out that surveyed a I assume lot of
00:06:25machine learning researchers all of
00:06:26these Alliance are people so and there’s
00:06:31a survey on the future progress on AI so
00:06:33what you think will happen when will
00:06:35happen and there’s a lot of really
00:06:37interesting things here that are amusing
00:06:40and possibly informative may be mostly
00:06:43amusing
00:06:45and let’s break it down on the easy end
00:06:47you see Angry Birds at roughly the same
00:06:49difficulty as the World Series of Poker
00:06:52that’s very unusual to me I thought
00:06:55Angry Birds was solved I’m pretty sure
00:06:56Angry Birds is solved and the World
00:06:58Series of Poker is actually sounds
00:07:00really hard but it’s it’s down there
00:07:02near the easy end on the difficult and I
00:07:06find it really interesting that AI
00:07:08researchers think it’s it’s like the
00:07:10second top there a researcher is
00:07:13significantly harder than math
00:07:15researcher seems like a 50 year gap
00:07:18between math research being solved an AI
00:07:20research being solved not saying that I
00:07:22agree or disagree it’s just interesting
00:07:24to point out it actually looks like the
00:07:27gap between AI researcher and math
00:07:29researcher is larger than the gap
00:07:31between math researcher and playing
00:07:33Angry Birds at a human level so I yeah I
00:07:38I don’t know if this reflects something
00:07:40about the field maybe that’s why there’s
00:07:41no good deep learning theory right now
00:07:43but who knows what’s going on but
00:07:46importantly for knowing what deep
00:07:48learning can do for us is there’s a lot
00:07:50of differing opinions on what’s going to
00:07:52happen and when it’s going to happen
00:07:53there’s some people who think that we
00:07:54will be getting generally AI in roughly
00:07:5710-15 years and there’s people who think
00:08:01it’s over a hundred years away I’m
00:08:03definitely in the latter camp to clarify
00:08:06but this makes it seem like answering
00:08:08the question what can I do for you
00:08:09correctly be really hard because there’s
00:08:12just so many different opinions how do
00:08:13you really know what the correct answer
00:08:14is it’s a tough problem but luckily I’m
00:08:17the only one on stage so I can just say
00:08:19whatever I want and no one can disagree
00:08:22with me I guess you can disagree with me
00:08:23in the Q&A and it’ll be a good debate
00:08:25but I’m gonna say what I think on this
00:08:29and take it with a grain of salt doesn’t
00:08:31reflect me it does reflect me it doesn’t
00:08:32reflect Google doesn’t reflect anything
00:08:34else I do like to stand on the shoulders
00:08:37of giants though and I think there’s
00:08:38been some people who have said things
00:08:40that resonates a lot with me and like
00:08:42really like when they say some things
00:08:44really precisely I feel like that helps
00:08:46refine my thinking on the problem this
00:08:49is one that I don’t know if I agree with
00:08:51but it’s a really strong statement and
00:08:52it seems like it could be a pretty good
00:08:54heuristic this is by and ruing who is no
00:08:56longer the chief scientist at Baidu
00:08:58but he says if a typical person can do a
00:09:01mental task with less than one second of
00:09:03thought we can probably automate it
00:09:05using AI either now or in the near
00:09:07future I can’t think of that many
00:09:10counter arguments that don’t require
00:09:12like a lot of like very specific domain
00:09:15knowledge like maybe people who play
00:09:16games a lot can play those games really
00:09:19fast cuz they’ve practiced it well but
00:09:21roughly it seems like a pretty good
00:09:23heuristic and it’s a very like strong
00:09:24statement so maybe this is something
00:09:27that could guide you answering that
00:09:28question another thing that isn’t as
00:09:32specific but I think is very important
00:09:36to consider is that a lot of the deep
00:09:39learning successes today have been I
00:09:42haven’t used the word simple but I feel
00:09:45like they’re more simple memorization
00:09:47problems and not really thinking
00:09:50problems it’s always hard to really say
00:09:54what this thinking really means because
00:09:55that might be a moving goalpost of like
00:09:57of course if algorithm is not thinking
00:10:00it’s using an a-star algorithm or
00:10:01something while we might think that’s
00:10:03kind of like thinking but even in this
00:10:06case it seems like when you’re when a
00:10:09task requires multiple steps of
00:10:11reasoning where you can’t like use
00:10:14heuristics to jump all the way from
00:10:16input to output it seems the deep
00:10:17learning has not been very good at that
00:10:19especially not without a lot of help
00:10:21which leads me to my general rule which
00:10:25is deep learning is an appropriate tool
00:10:27for supervised direct pattern matching
00:10:30tasks bonus points if you can design
00:10:32priors that are particularly suited to
00:10:34your problem the the priors in this case
00:10:36are specific layers that are popular for
00:10:39certain tasks but we don’t have to get
00:10:41into that right now
00:10:42even though feel free to ask me in the
00:10:44Q&A but here when I say supervised I
00:10:47mean that we tell the model directly
00:10:49what the correct answer is so roughly a
00:10:51human or some other process figures out
00:10:54what the right answer is via some means
00:10:56and tells the model this is what you
00:10:59should be outputting next time there are
00:11:01there have been incredible successes
00:11:04using reinforcement learning which is
00:11:07not supervised especially in the
00:11:10game-playing domain so if you’ve seen
00:11:12deep Minds deep queue networks playing
00:11:15Atari or deep Minds alphago playing go
00:11:18those use a lot of reinforcement
00:11:21learning and if they definitely have in
00:11:24some successes but how I feel about
00:11:26reinforcement learning is that it can
00:11:27work but you don’t want to rely on it
00:11:29working and in the bay everyone hasn’t
00:11:32started up on everything and there’s
00:11:33been a lot of people who kind of have
00:11:35bet their companies on this is a
00:11:37reinforcement learning problem let us
00:11:39sell people on using deep reinforcement
00:11:42learning to get this working and ending
00:11:43up with vaporware and kind of a sad end
00:11:49to that supervise for it but as far as
00:11:51direct pattern matching goes this goes
00:11:52back to what I was saying earlier where
00:11:54you want simple relationships between
00:11:56the input and the output
00:11:58almost almost like that a fraction of
00:12:02the input directly maps to some fraction
00:12:04of the output in some sort of additive
00:12:07ish way it doesn’t have to be completely
00:12:09additive but usually having some easy
00:12:12mapping allows you to bootstrap the more
00:12:13complicated mappings and a lot of the
00:12:17more complicated mappings turn out to be
00:12:20like lots of little simple mappings
00:12:22composed together and this kind of thing
00:12:24seems to be how deep learning tel tends
00:12:26to work this is all very vague but I am
00:12:29about to talk about some specifics about
00:12:31where deep learning has succeeded and
00:12:33where it seems have not succeeded yet
00:12:36another disclaimer this is only a subset
00:12:39of the potential cool things to talk
00:12:40about and I’m only talking about the
00:12:44intersection of things I find
00:12:45interesting because I want the slides to
00:12:47be interesting there’s lots of like
00:12:48little things that are cool but maybe
00:12:50wouldn’t be that interesting to people
00:12:51and visual and things that I know how to
00:12:54put in a presentation so I I have some
00:12:57attempts at videos but they’re optional
00:12:58but I there’s some cool lots of cool
00:13:01working audio but I just have no idea
00:13:03how to put that in a presentation and
00:13:05[Music]
00:13:06yeah maybe that’s my bad but we can
00:13:09solve go and build robots but technology
00:13:11isn’t there yet for reliable audio and
00:13:14video this is also what happens when I
00:13:16make the graphics myself so all of the
00:13:19pretty animated graphics have been
00:13:20stolen from other Googlers who know how
00:13:23to do ours
00:13:25back on topic let’s start with the
00:13:28easiest thing whenever you have a metric
00:13:30that when that metric goes up money goes
00:13:33up you probably want to use machine
00:13:36learning possibly deep learning but
00:13:38definitely machine learning this is
00:13:40actually I would describe the main use
00:13:41case of machine learning
00:13:42I have knobs to turn some combinations
00:13:45of these knobs are better than others
00:13:46how do I turn them basic stuff but we’re
00:13:49saying a big thing that people get
00:13:53caught on is unsupervised learning it’s
00:13:57a very interesting research problem but
00:14:00if you want to do anything practical I
00:14:02would probably advise you to not do that
00:14:05I think this is actually absolutely
00:14:07excellent advice rather than trying to
00:14:09if you can spend a month figuring it out
00:14:12on supervised learning please do it that
00:14:14that will solve a lot of people a lot of
00:14:15time if you could spend a year that
00:14:18would probably save a lot of people out
00:14:19of time if you could do ten years you’re
00:14:21probably on track with the rest of the
00:14:22field so if you have a problem that you
00:14:26care about don’t try to do some magic
00:14:29where you don’t know if it’s gonna work
00:14:30label some data usually these things are
00:14:32a lot more data efficient and people say
00:14:34they are and sticking to supervised
00:14:36learning will be much easier for your
00:14:38sanity as well as your eventual impact
00:14:42speech recognition has done really well
00:14:45really really well people think that
00:14:47this is probably going to be one of the
00:14:49biggest changes to interfaces in not
00:14:53just our lifetimes but in the next
00:14:54decade right now people don’t like to
00:14:56talk in phones because they could be
00:14:58kind of sucks but people can talk much
00:15:01faster they can type and a lot of people
00:15:02don’t know how to type very well so this
00:15:04could completely change the way people
00:15:05interact with electronics things like
00:15:08Google glass or I hear it’s really big
00:15:11in China speech recognition there’s all
00:15:13sorts of things that this could enable
00:15:14and this is only a fraction of the cool
00:15:16things happening in audio but I not
00:15:19think to talk about that much but
00:15:20there’s things with generating
00:15:22generating audio generating music lots
00:15:25of cool stuff there translation this
00:15:29animation is really cool and this
00:15:31problem is really cool this is showing
00:15:33that not only can deep networks improve
00:15:36on like the traditional statistical mess
00:15:39that things like Google Translate used
00:15:41to do but where you just have matching
00:15:44corpuses or Corp I I don’t know what the
00:15:46plural is but you can also translate
00:15:49between language pairs that you’ve never
00:15:51you don’t even have matching corpuses on
00:15:53so in this example you have English to
00:15:56Japanese pairs as well as Japanese so
00:15:59sorry English to Japanese in English to
00:16:00Korean and using these networks you can
00:16:03actually translate directly between
00:16:04Korean and Japanese without ever seeing
00:16:07paired data between Korean and Japanese
00:16:09which is actually huge
00:16:11it could enable a lot of translation on
00:16:14languages that between languages that
00:16:16there’s just no data on and you can do
00:16:18it in a much more accurate way because
00:16:20you don’t need to translate into an
00:16:21intermediate language where you lose
00:16:23some information if you ever like do
00:16:26Google like what’s the game where you
00:16:29have a like a Markov chain with Google
00:16:30Translate you start with a thing you
00:16:32translate to one language you translate
00:16:33back eventually becomes garbage and
00:16:35nothing like the original thing you said
00:16:37and you just avoid that problem entirely
00:16:39with this image classification this is
00:16:43like the bread and butter of deep
00:16:44learning it’s what made deep learning a
00:16:46big deal people it was kind of a not
00:16:49mainstream thing until about two twenty
00:16:52twelve when deep learning one this
00:16:54imagenet competition and beat all of the
00:16:58other things by a fairly large margin
00:17:00and made everyone realize hey this
00:17:02solves problems that nothing else could
00:17:04solve before and there’s real-world
00:17:07applications to this like google photos
00:17:10an example there’s a lot of api’s where
00:17:12people have made a business of telling
00:17:16you what’s in an image people do face
00:17:18classification Faith’s
00:17:20detection there’s a lot of money and
00:17:24sentiment recognition you know like have
00:17:27a camera here and look at the room tell
00:17:28them if they’re enjoying the talk or not
00:17:30based on like people’s smiles and stuff
00:17:32maybe not for talks but like for ads and
00:17:35stuff something that can’t be done yet
00:17:38though is unbiased image classification
00:17:41or it’s still a lot of work this was a
00:17:43huge issue for Google photos actually
00:17:47like I think it’s like a few days after
00:17:49they released it people were complaining
00:17:51on Twitter that
00:17:53their friends were being classified as
00:17:58gorillas due to a lack of diversity in
00:18:01the training data and this is kind of
00:18:03unavoidable when you have imperfect
00:18:05datasets I actually don’t know how they
00:18:07solve this they might have just removed
00:18:09some of the classes that could be been
00:18:11taken as offensive but that’s just a
00:18:15hack right like we want like real
00:18:17algorithms that don’t make these kinds
00:18:19of stupid mistakes talking about not
00:18:22making stupid mistakes a problem near
00:18:25and dear to my heart is medical imaging
00:18:26they’ve been a bunch of huge successes
00:18:29on medical imaging in particular there’s
00:18:32been some really cool stuff done reading
00:18:34x-rays and CT scans cool stuff with
00:18:36segmenting pathology scans detecting
00:18:40diabetic retinopathy all of these things
00:18:43it’s people have been getting superhuman
00:18:46results like better than what seems to
00:18:48be the best doctors and hopefully very
00:18:51soon this kind of stuff will be like
00:18:52reaching the end users and helping
00:18:54people so this is a really exciting area
00:18:56of deep learning progress similar in
00:19:00that vein it’s not limited to either 2d
00:19:04images or having a single prediction per
00:19:06image you can do what’s called semantic
00:19:09segmentation where you label each pixel
00:19:11or in this case voxel in an image and
00:19:14you can also it also works for high
00:19:16dimensional data so this for example is
00:19:183d segmentation of I believe in neuron
00:19:21and this algorithm actually is iterative
00:19:23and how it like expands over time and
00:19:25this is very similar to how a human
00:19:28would segment a neuron it would not just
00:19:30say all those months here’s a neuron if
00:19:32it starts at something and being like
00:19:34okay this is close to this other thing
00:19:35this is maybe a neuron so we are as we
00:19:39were like expanding the reach of deep
00:19:41learning more people are designing more
00:19:42and more of these priors to build into
00:19:44the architectures to do much smarter
00:19:46things so whenever I say not yet on
00:19:48something it might be that technology is
00:19:50there P we just haven’t tried hard
00:19:52enough talking about and not yet there’s
00:19:56been some really cool work on image
00:19:58captioning so instead of given an image
00:20:00output a object in the image it’s given
00:20:04an image describe the image
00:20:06and this is a much harder task because
00:20:08there’s a lot of things that can go on
00:20:10in an image and there’s a lot of
00:20:11possible ways to describe an image so
00:20:13how do you say something is riot and
00:20:15what what set of things do you choose to
00:20:19have something described and this is
00:20:21pretty good like these descriptions are
00:20:24actually this is a good case there’s
00:20:27many bad cases of this but they still do
00:20:29make some really dumb mistakes it might
00:20:31reflect underlying issues with our
00:20:33imaging models or it might be due to
00:20:35dataset size but this is still an open
00:20:36research problem similarly to that it’s
00:20:41not very good at answering questions
00:20:42about images or stories it can be good
00:20:47at finding specific things in the images
00:20:49but there’s other things that seem to be
00:20:52easier than finding a thing or just as
00:20:54easy as finding a thing that deepening
00:20:56currently it’s not good at like counting
00:20:58if you ask good this is I don’t have a
00:21:00counting example here but if you have
00:21:02like a bowl of oranges and you ask like
00:21:04how many oranges are in this bowl this
00:21:06sounds like a very easy task but it’s
00:21:09quite hard for models right now so
00:21:12that’s a big problem talking about big
00:21:16problems we definitely are nowhere close
00:21:18to automating research this is a great
00:21:22tweet we’re the researchers were the
00:21:26ones that wanted to make the AI do all
00:21:27the work and play games and while they
00:21:29play games but instead it’s the opposite
00:21:31right the yeah is just playing games all
00:21:33day and researchers are working harder
00:21:35than ever it’s a tough life I think the
00:21:38comments on this were equally great
00:21:40because maybe this is a sign that the AI
00:21:43is actually intelligent you know maybe
00:21:45it’s like just pretending to be dumb and
00:21:46being like why would I want to do all
00:21:48the work I’m just gonna keep playing
00:21:49games all day some aspects of research
00:21:53might be automated something that some
00:21:56people consider to be either boring or a
00:21:59waste of time or hard is designing these
00:22:02architectures in the first place and
00:22:03there has been some work in using deep
00:22:07learning to automate the architect the
00:22:10design of architectures for more deep
00:22:12learning and you get like these crazy
00:22:14things that no one would ever design
00:22:17yeah I would definitely not
00:22:20think to do that in the right so this
00:22:24stuff has had some fairly promising
00:22:27results I I put this under a maybe of
00:22:29what can be plausible because it’s both
00:22:32it was both very expensive and not quite
00:22:34as good as a state of the art but this
00:22:36seems like a really promising Avenue and
00:22:38a potential place that it could make a
00:22:41big impact so maybe all of our learning
00:22:44about architecture and studying this and
00:22:45trial and error maybe all of this will
00:22:47be outsourced to you know farms of
00:22:50computers somewhere and we could just
00:22:52you know stick to the high level tasks
00:22:54but life is rarely that kind a despite
00:23:00fake news to the contrary we are long
00:23:02ways away from automating software
00:23:04development there were some articles on
00:23:07algorithms automating coding and III
00:23:13think that some people were a little bit
00:23:16panicked on this maybe all of the
00:23:18articles when deepening automates X
00:23:20causes some panic but I hang around I
00:23:22hang out with lots of software engineers
00:23:24so they were worried for like a second
00:23:28until they realized that this thing was
00:23:30actually really really dumb not that the
00:23:32the work was done but how the algorithm
00:23:35did it was nowhere close to software
00:23:37engineering it was a slightly better
00:23:40heuristic for picking random bits of
00:23:43code together and doing trial and error
00:23:46on that code and as we all know that is
00:23:49absolutely not how we do software
00:23:51engineering right like we design stuff
00:23:53upfront not trial and error it’s like
00:23:55all done by the books yeah this
00:23:58algorithm definitely can’t do that
00:24:00so we our jobs are safe right guys
00:24:05ok so some people some people know what
00:24:08I’m talking about some great memes this
00:24:13is not really model output but if you do
00:24:15follow the field people love to have fun
00:24:19things in there some people actually
00:24:21wrote a conf a fait paper I think that’s
00:24:25pretty incredible it’s someone can like
00:24:27dedicate a research paper with I assume
00:24:30a real idea I didn’t read this but I
00:24:32assume it’s a real idea
00:24:33to a troll name I think it’s great and
00:24:36it shows like the speed of publishing in
00:24:38the field common Silicon Valley problem
00:24:43no deep learning will not solve all of
00:24:46your problems especially not your
00:24:48product definition problems it won’t
00:24:49find something useful for you to do and
00:24:52it will not make you magically rich
00:24:54despite a lot of belief to the contrary
00:24:57and similar to this image general chat
00:25:02bots are actually quite difficult you’d
00:25:06think that you just give you no model a
00:25:08dataset of two people talking and it’ll
00:25:10be able to replicate those people
00:25:12talking but it turns out that our
00:25:14language models are quite good at making
00:25:16things that look grammatically correct
00:25:18but are semantically quite terrible so
00:25:21they don’t have like they don’t have any
00:25:23history involved they don’t have there’s
00:25:25lots of issues of them and this like a
00:25:30misunderstanding too that led to a lot
00:25:32of companies starting products that
00:25:34ended up pivoting away from using deep
00:25:37learning at all and ended up using like
00:25:39an army of workers in the Philippines
00:25:42this manually doing the chatting for
00:25:44them which turns out to be a pretty
00:25:46economical way to do things but specific
00:25:50chat bots are very doable so when if you
00:25:53turn the problem from hey let’s generate
00:25:56arbitrary text to hey let’s pick among a
00:26:00small set of valid responses things
00:26:03become a lot easier this is in boxes
00:26:07smart reply which apparently is used by
00:26:09over ten percent of mobile infox replies
00:26:13which sounds like a lot of
00:26:15qualifications but I just think it’s
00:26:18cool that something started out as an
00:26:19April Fool’s Day is now real April
00:26:23Fool’s Day joke it’s also sweet
00:26:25animation but this kind of stuff is very
00:26:27plausible and I think that people who do
00:26:29use machine learning for chat BOTS will
00:26:31end up constraining the problem quite a
00:26:33bit and that’s actually very doable if
00:26:35you’re trying to classify like do I have
00:26:37enough information or is this person
00:26:39satisfied or do I need to pull another
00:26:42human in to actually chat through this
00:26:44person that’s much more doable than hey
00:26:47automatically solve this person’s IT
00:26:50issues which sounds really hard and
00:26:55similar in vain to that coherent text is
00:26:59quite a challenge like long any long
00:27:01amount of text a lot of journalists
00:27:03journalists I feel like a big victim to
00:27:05hype because it’s kind of their fault
00:27:07uh-huh and they’re kind of worried about
00:27:10their jobs about RDR deep nets and like
00:27:13start writing articles for us and the
00:27:17answer seems to be no so if you’re a
00:27:19journalist
00:27:20don’t worry sorry
00:27:32did you say it’s hard to investigate
00:27:34journalists as an AI
00:27:41I couldn’t quite hear the last part of
00:27:43that oh yeah for sure
00:27:53he said as may I it’s hard to do
00:27:55investigative journalism that is
00:27:58definitely true debatable about how much
00:28:02investigative journalism current
00:28:04journalists do but that’s definitely the
00:28:08case I think in this case it is even the
00:28:14worry that a lot of journalism is read
00:28:17stuff on Twitter turn it into an article
00:28:19hope to get lots of clicks make click
00:28:22vadie headline topic definitely not all
00:28:26of it but some fraction of it is that
00:28:29and I think that there is some worry
00:28:30about this like I believe in finance
00:28:33there’s a big race to like who can
00:28:35publish these articles first based on
00:28:38various data sources and if you’re not
00:28:42talking about quality but speed these
00:28:45things definitely have a speed advantage
00:28:48yeah III I’m not I’m personally not
00:28:51worried about journalists jobs being
00:28:53taken for sure cool sorry sorry I
00:29:07couldn’t hear you very well but thank
00:29:09you for yelling that time yeah so
00:29:15something that seems to be really
00:29:17promising I actually think that this is
00:29:18one of the most promising upcoming uses
00:29:21of deep learning that’s like not quite
00:29:23there but might be there and it’s like a
00:29:26would be a really sexy field to get into
00:29:27with robotics it seems like there’s a
00:29:29lot of really good stuff happening with
00:29:31imitation learning and a lot of people
00:29:34are invested just working oh it is
00:29:36working this is pretty cool people are
00:29:39investing in a lot of the the research
00:29:42labs are investing and getting the
00:29:47robots training together in like how do
00:29:49we collect lots of data for robots to
00:29:51get them working automatically because
00:29:52right now
00:29:53at least to my understanding I’m no
00:29:55roboticist is that majority of the work
00:29:57done by robots is done manually and if
00:30:00we can like make it a lot easier to
00:30:02train robots to do things that we care
00:30:03about maybe
00:30:04all of a sudden we’re going to have like
00:30:05more general programmable robots that
00:30:07people can do stuff with so that I think
00:30:11is really really promising and at least
00:30:15from a research perspective of someone
00:30:16who reads the papers and like keeps up
00:30:18with what people are doing it seems very
00:30:21plausible that this kind of thing could
00:30:25make a breakthrough in the near term
00:30:26especially with what’s called imitation
00:30:29learning where robots rather than
00:30:31learning by trial and error which can be
00:30:33very hard they just learn to copy humans
00:30:35which is goes back into the rule of
00:30:37thumb I was talking about we’re giving
00:30:41giving these algorithms supervised data
00:30:43telling them what to do generally works
00:30:44a lot better than hoping for magic that
00:30:49you know hoping that they will magically
00:30:51figure out the thing to do which is what
00:30:53a lot of the field is trying to get
00:30:55working right now depending on who you
00:30:58ask
00:30:59game playing I would count categories
00:31:02that is not yet there there have been
00:31:05some amazing successes in game playing
00:31:07but a lot of those successes aren’t
00:31:10quite super general a lot of it it’s
00:31:12like very input simple input output
00:31:14mapping like I was mentioning so Atari
00:31:17seemed to be a lot of that debatable
00:31:20whether or not go was that even though
00:31:21that was definitely a huge win but
00:31:23there’s been other games where models
00:31:26are nowhere near as successful so things
00:31:29like even like very simple Minecraft
00:31:31mazes it’s still the model still aren’t
00:31:35quite there yet or recently there’s been
00:31:37a bunch of work on doom visual doom like
00:31:40do them from the pixels and this model
00:31:44actually super cool let me see here does
00:31:47this work you can skip to the fighting
00:31:52so there’s been a lot of progress on
00:31:54that really recently this is like the
00:31:57what was the state of the art in 2013 if
00:32:00you can tell it’s like pretty dumb like
00:32:02shooting a wall right now let’s see here
00:32:07yeah this is so this is a little bit
00:32:09smarter this was state of the art in I
00:32:12would say 2016 ish mid 2016 this is
00:32:18still pretty dumb
00:32:19and people been making a lot more
00:32:22progress recently with this we’re look
00:32:25at this this is actually intimidating
00:32:28it’s like moving around its shooting
00:32:31intelligently etc etc there’s been a lot
00:32:35more progress being made in this and it
00:32:36seems that we’re nowhere near close to
00:32:39or at least to my knowledge solving
00:32:42something like Starcraft but it’s really
00:32:45promising and people are putting a lot
00:32:46of effort into this so the likelihood
00:32:48that we make some big breakthroughs in
00:32:50the coming years seems to be likely cool
00:32:55and this is category of stuff it’s like
00:32:58what I am one of the most excited about
00:33:01just because I would never consider
00:33:03using like these extremely powerful
00:33:06classification models for artsy things
00:33:08maybe that’s just me but some of these
00:33:12use cases are incredibly creative and
00:33:14incredibly cool and like this stuff is
00:33:17amazing this one came out pretty
00:33:19recently and it learns to transform
00:33:22images from different domains so
00:33:25transforming like a zebra into a horse
00:33:27or vice versa so image transformation it
00:33:31can be done can even be done with videos
00:33:34so this is like actually really done by
00:33:37a model it’s not like cherry pick data
00:33:42but so it’s actually a transform this
00:33:48video and like this is not seamless but
00:33:50that’s pretty good better than I could
00:33:53do with Photoshop which is not saying
00:33:55much but like this is like pretty
00:33:56impressive and I would not even have
00:33:57thought of this as a use case like hey
00:33:59I’m a machining researcher at Google I
00:34:03have a you know giant cluster
00:34:05I’m gonna transform a horse into a zebra
00:34:07right
00:34:10unfortunately this kind of thing is not
00:34:13completely reliable but this is pretty
00:34:18amusing so like with all machine
00:34:21learning like making it completely
00:34:23reliable can be challenging talking
00:34:26about that there’s been an app that’s
00:34:28been gaining popularity called face app
00:34:31that
00:34:32does facial transformation so in the top
00:34:34left you see the original photo top
00:34:37right you see like a more manly
00:34:40transformation you know like more edgy
00:34:43chin bill beardy then bottom left and
00:34:47old miss transformation and bottom right
00:34:50a smiling transformation and this is
00:34:53actually pretty good pretty good and
00:34:56like an app can do it on your phone no
00:34:58human input it just does it it’s pretty
00:35:01impressive that you can do this and this
00:35:03is a really cool use case unfortunately
00:35:05it’s not perfect in particular it’s also
00:35:08suffers from that bias problem like with
00:35:11many other things when you turn a cool
00:35:13model into a product it there’s a
00:35:14different set of requirements in this
00:35:16case they had a transformation which
00:35:19makes a person’s face hotter and one of
00:35:23the things it did was it always lighten
00:35:25skin which was offensive to some people
00:35:28yeah that they had to pull that feature
00:35:31I think or change I think they actually
00:35:33changed the name from hot to something
00:35:35else I can’t remember art is doable this
00:35:41is art from scratch or unconditional art
00:35:43like it’s like these models can just
00:35:46create these artsy things and I think
00:35:50that this is really really cool I
00:35:52personally think that these both look
00:35:54really good can I get a show of hands of
00:35:58who thinks the one on the left is better
00:36:01what about the right oh it looks like a
00:36:05tie I made the one on the left so I was
00:36:09hoping that people would vote for that
00:36:11one cool I think they’re both really
00:36:15cool I would definitely have a poster of
00:36:17that in my room or a painting of that in
00:36:19my house then this kind of thing like
00:36:22who would have even thought that as a
00:36:24side effect of these really powerful
00:36:27actually useful things we would get art
00:36:46oh yeah I’m definitely not claiming that
00:36:49this is the like the the the first thing
00:36:53in terms of algorithmic arts it’s just I
00:36:56find to just be a really cool use case
00:36:58of deep learning because when III don’t
00:37:02think ten years ago people would have
00:37:04imagined like yeah imagine all of these
00:37:06cool pictures will make and every
00:37:09actually every time there’s a new use
00:37:10case in art I’m just amazed like who
00:37:13thought of this like who spent their
00:37:15time on this and I’m thankful for that
00:37:17because I wouldn’t have done it but I
00:37:19think it’s really awesome and I think in
00:37:22some ways it’s also kind of cool that
00:37:26unlike fractals or something like that
00:37:29it feels like there’s there’s a lot more
00:37:32like there’s more unknown unknowns in
00:37:34this right now which makes it really
00:37:36promising as well maybe that’s from my
00:37:39misunderstanding of art though or
00:37:41algorithms or anything I’m not expert in
00:37:44any of this stuff sketching with another
00:37:50recent use case where you just train in
00:37:52a data set of humans drawing little
00:37:53things and the things in the the top
00:37:56corner up there is things that the model
00:37:58drew and in the bottom here you can
00:38:01actually do math on sketches so you take
00:38:04like a cat face you add in a pig with a
00:38:07body you subtract a pig face and you end
00:38:10up with like a cat with a body and like
00:38:13it’s kind of cool that it works I mean
00:38:18the math checks out so awesome for for
00:38:26the non artists in the room you can also
00:38:29turn what is arguably not art in the
00:38:33bottom-left into something that is
00:38:35potentially art so this is another very
00:38:38cool use case where you like it can like
00:38:41enable people to it becomes almost like
00:38:46you know a new artistic medium right
00:38:49where you can now use these things to
00:38:51enhance existing art to do things maybe
00:38:53that people wouldn’t have done before or
00:38:55enable people who couldn’t have done
00:38:57this before
00:38:57or maybe just make it more faster
00:38:59thing like that it feels almost like you
00:39:04know like a new instrument from the
00:39:06musical sense so this stuff is really
00:39:08cool style transfer I think this is
00:39:11crazy because a year and a half ago this
00:39:15was already looking really good and it’s
00:39:18just gotten better and better so this is
00:39:19like going so well I actually I should
00:39:24have put the old pictures here as well
00:39:25but like these are the new what I think
00:39:28is the latest in style transfer and this
00:39:31is pretty good like you can see
00:39:34transferring the style of a fire into a
00:39:37bottle like this is like a professional
00:39:39Photoshop job and I guess and this is
00:39:44like impressive and I would want to do
00:39:46this and I look forward to this being
00:39:49able to be done for me because I don’t
00:39:51want to implement it myself but there’s
00:39:55a lot of really cool stuff being done
00:39:57with style transfer and this stuff is
00:40:00really pragmatic because like
00:40:02aesthetically this is already like very
00:40:04high quality this is my crowning
00:40:08achievement actually mixing my face with
00:40:11that of a Pokemon probably my best
00:40:14achievement and deep learning definitely
00:40:16works would recommend trying it again
00:40:18and probably newer stuff will work even
00:40:20better and as far as specific go there’s
00:40:25all sorts of other things a rough
00:40:27formula is taken input that is similar
00:40:31to another input that deep learning has
00:40:33succeeded on like images audio raw text
00:40:37other domains like that pick a response
00:40:40that is a relatively simple mapping from
00:40:44that input so nothing too complicated
00:40:47but simple mappings like are they human
00:40:51faces and collected data set trainer
00:40:54model usually something like that gets
00:40:58just something that works quite well and
00:41:01as far as what it can do if you if you
00:41:05pick the right things generally it
00:41:07generally makes it is the valgar isn’t
00:41:11helped you a lot in doing
00:41:13a lot of the easy work for you getting
00:41:15like the last bit of presents always a
00:41:17lot of work but you you’ll know if you
00:41:21can get it which is it which makes it a
00:41:23little bit easier
00:41:24oh so back to the big questions how will
00:41:31the world change I think this is a great
00:41:34tweet like Andrew Inge
00:41:36I do believe automation and steroids is
00:41:38the right way to think about it not
00:41:39sentience or AI overlords or anything
00:41:43else like that I actually would be quite
00:41:47pleasantly surprised to see generally I
00:41:49in my lifetime just because I think
00:41:52that’s so unlikely and that’s not
00:41:54because despite what my current pants
00:41:56might imply that I’m one of the live
00:41:58fast die young types it’s it’s that I I
00:42:02think that it’s quite quite a ways away
00:42:04though I would love to be wrong as far
00:42:09as how the world change I won’t claim to
00:42:12be an expert on the societal effects of
00:42:14automation but luckily this guy would he
00:42:16had a TED talk called will automation
00:42:19take away all our jobs all seems like a
00:42:21little bit of a weasel word here it has
00:42:23over a million views and bet riches law
00:42:27of headlines applies here any headline
00:42:30that ends in a question mark can be
00:42:31answered by no so the answer is no
00:42:35saving you 18 minutes the claim is that
00:42:39AI automation just like other automation
00:42:41will take some jobs away but it will
00:42:43probably do much more transformation of
00:42:45jobs because a lot of jobs aren’t these
00:42:47just simple mappings and there’s more
00:42:49complicated new instincts to them but
00:42:53automation increases the leverage of any
00:42:55person and there will be a lot more jobs
00:42:58that we just can’t imagine will happen
00:42:59so that’s his view I don’t have strong
00:43:05views in this yeah what can deep
00:43:10learning do the fields really exciting
00:43:12there’s a whole lot of things that we
00:43:13can do now that we previously couldn’t a
00:43:15lot of things that were once thought to
00:43:16be really really hard are now doable
00:43:19people fought out of solving go was like
00:43:21a hundred years away or more and there’s
00:43:25all sorts of fields that this could
00:43:27affect
00:43:27as far as specifics ago which is what
00:43:30would actually be more useful to you
00:43:32guys
00:43:33the answer is it’s kind of complicated
00:43:36my advice would be to look at example
00:43:38failure and success cases network with
00:43:41researchers and people in industry or
00:43:42use someone’s rule of thumb maybe my own
00:43:45maybe take it and change it but really
00:43:48you want to build your own mental
00:43:49classifier of what isn’t isn’t possible
00:43:52and refine that classifier around the
00:43:54set of problems that you specifically
00:43:56care about so if you want like a
00:43:58specific like I want to figure out if
00:44:00deep learning can find X in a genomic
00:44:03data set you like go into the research
00:44:05look at that figure out like what seems
00:44:07possible what’s not and there’s so many
00:44:10problems out there that you might have
00:44:12you might end up being the world expert
00:44:14in knowing if deep learning works for
00:44:16your problem so it’s just there’s so
00:44:18much opportunity right there that if you
00:44:21ask anyone there probably will tell you
00:44:22an answer because people love to give
00:44:24answers but they probably won’t give you
00:44:25a very good one including myself so
00:44:28right now it’s unavoidable to do some of
00:44:30that work unless you do something that
00:44:32someone’s already solved but that’s kind
00:44:34of a cop-out answer the last big
00:44:38question it was how do I take advantage
00:44:40of these trends I think it’s a lot like
00:44:42learning softer engineering especially
00:44:44back in like when the internet was young
00:44:47and my answers scratch your own itch
00:44:49play around with it a lot of the work on
00:44:52art specifically was done by hobbyists
00:44:54and not researchers and we have no idea
00:44:57yet what can be done on you know the
00:44:59problems you care about and for all you
00:45:01know you might be sitting on depending
00:45:03the next killer app that no one else is
00:45:04thought of and it’s scratching your own
00:45:07itch leads to something valid for others
00:45:09start a company on it there’s lots of
00:45:12money for companies going around right
00:45:14right now and the world needs more AI
00:45:16companies that actually provide value
00:45:18I’m not gonna name names and also
00:45:23prepare for like a sweet transition
00:45:26consider joining Google it’s the best or
00:45:30any other like AI focused company which
00:45:32is like interesting impactful problems
00:45:34and the resources to solve those
00:45:36problems
00:45:36thank you
00:45:39you