Press "Enter" to skip to content

GOTO 2017 • Deep Learning: What It Is & What It Can Do For You • Diogo Moitinho de Almeida


[Music]
cool I’ll get started I’m Diego I work
at Google brain but that’s all I can
tell you my normal rule is if I told you
I’d have to kill you but there’s a whole
lot of you guys which makes that kind of
impractical right now a standard
disclaimer everything I say reflects my
own opinions it’s not representative of
my employer I have not been at Google
long enough to know any secrets though I
will admit that I’ve pilfered some
publicly available slides so if there is
a Google bias it’s not because they made
me do it it’s because I’m lazy as far as
some background for myself
I broke a 13-year losing streak for the
Philippines in the International math
Olympiad I got the top prize in the
world in the inter display compass in
modeling and if you’re familiar with
kayo competitions I also want one of
those so I like to compete and do all of
these kinds of things and hopefully that
convinces you that I know that I’m
talking about but let’s get started the
presentation is deep learning what is it
and what can it do for you
but I think the very first question is
why should I care and that reminds me of
a story a machine learning researcher a
cryptocurrency expert and an Erlang
programmer walk into a bar
Facebook buys the bar for twenty seven
billion dollars and also another
disclaimer you may not know this but I’m
both a machine learning researcher and
from San Francisco that means all of my
information comes from Twitter that’s
not a joke so prepare for that for my
slides but back to why you should care
machine learning artificial intelligence
deep learning they’re all getting a lot
of press these days they’re all doing
lots of stuff and there’s lots of hype
lots of news articles about all sorts of
things and everyone seems to want to get
into it but people don’t really know
what they’re talking about it seems the
only thing everyone’s really sure of is
that artificial intelligence and it
seems like very recently in particular
deep learning will be a catalyst for a
lot of change that’s happening and
people are asking any questions all the
time about this kind of thing and some
of these recurring themes are how will
the world change
what can a AI and in particular deep
learning do and how do I take advantage
of these trends
even the legendary programmer Jeff Dean
has said if you’re not considering how
to use deep neural nets to solve your
problems you almost certainly should be
it almost sounds like a threat either
way I hope you’re motivated to learn
that’s all I have for motivation so
let’s get started into what is deep
learning and it’s quite easy this is
deep learning you could memorize this
diagram this will make the presentation
a lot easier so that’s basically it just
kidding this is neither complete there’s
a lot more to it than that and it’s also
pretty complicated we’re gonna start
with something much more simple namely
calculus may not sound right it’s even
chapter 1 of the book calculus made easy
it’s titled to deliver you from the
preliminary terrors but we only need a
little bit of calculus and particularly
we need an algorithm called gradient
descent the derivative of one thing with
calculus tells you is how to take
derivatives and a derivative loosely
speaking if it tells you how a functions
output changes when you change its input
and gradient descent is just moving
along the direction of the derivative in
order to minimize a function that you
can take the derivative of so the
insight into all of machinery Sigma Xin
learning is figure out how to frame your
problem in such a way that what you care
about is differentiable or a proxy of
what you care about is differentiable
and then minimize it and this like one
extremely simple equation summarizes
almost all of the recent work in machine
learning that’s happened in the last
half decade roughly there’s obviously
exceptions to this rule but majority of
what’s done either has been done with
this exceptionally simple thing or can
be done with this exceptionally simple
simple thing and that’s basically it as
far as what deep learning really is in
its essence
this might sound well good but this is
just machine learning when’s it become
deep learning also easy it’s when you
make it really deep it might sound like
a joke but acts
really what happens when you have these
you know these machine learning used to
be composed these very very simple
functions because this is all we knew
how to optimize and what happens when
you stack multiple of these simple
functions together you get something
that’s much much more powerful if we
don’t really know how much more powerful
it is some might claim it’s
exponentially more powerful but either
way we know it’s much more powerful and
simply stacking these things and using
the simple algorithm is what’s caused
the deep learning revolution to hit and
it’s just using same old simple
algorithm even though as a caveat to
that we part of the hard part of deep
learning is knowing that the simple
algorithm will work for these very
complicated models that have like stacks
of layers and making problems non convex
is that it for deep learning yes this is
it roughly I’m skipping all sorts of
details that I’m sure will be covered
later from software that makes things
easier to write like tensorflow or pi
torch to hardware that makes things
faster to run like GPUs or multi-core
CPUs or TP use to commonly use some
functions or layers that people have
done lots of trial and error and just
found to work well for some of today’s
problems without very good justification
and also commonly use combinations of
these layers or architectures that
similarly people have used trial and
error and found to work without very
good justification but for the purpose
of that talk the simpler algorithm is it
next big question is what can it do for
you
this may there was a recent paper that
came out that makes it a little bit
easier to answer because this article
came out that surveyed a I assume lot of
machine learning researchers all of
these Alliance are people so and there’s
a survey on the future progress on AI so
what you think will happen when will
happen and there’s a lot of really
interesting things here that are amusing
and possibly informative may be mostly
amusing
and let’s break it down on the easy end
you see Angry Birds at roughly the same
difficulty as the World Series of Poker
that’s very unusual to me I thought
Angry Birds was solved I’m pretty sure
Angry Birds is solved and the World
Series of Poker is actually sounds
really hard but it’s it’s down there
near the easy end on the difficult and I
find it really interesting that AI
researchers think it’s it’s like the
second top there a researcher is
significantly harder than math
researcher seems like a 50 year gap
between math research being solved an AI
research being solved not saying that I
agree or disagree it’s just interesting
to point out it actually looks like the
gap between AI researcher and math
researcher is larger than the gap
between math researcher and playing
Angry Birds at a human level so I yeah I
I don’t know if this reflects something
about the field maybe that’s why there’s
no good deep learning theory right now
but who knows what’s going on but
importantly for knowing what deep
learning can do for us is there’s a lot
of differing opinions on what’s going to
happen and when it’s going to happen
there’s some people who think that we
will be getting generally AI in roughly
10-15 years and there’s people who think
it’s over a hundred years away I’m
definitely in the latter camp to clarify
but this makes it seem like answering
the question what can I do for you
correctly be really hard because there’s
just so many different opinions how do
you really know what the correct answer
is it’s a tough problem but luckily I’m
the only one on stage so I can just say
whatever I want and no one can disagree
with me I guess you can disagree with me
in the Q&A and it’ll be a good debate
but I’m gonna say what I think on this
and take it with a grain of salt doesn’t
reflect me it does reflect me it doesn’t
reflect Google doesn’t reflect anything
else I do like to stand on the shoulders
of giants though and I think there’s
been some people who have said things
that resonates a lot with me and like
really like when they say some things
really precisely I feel like that helps
refine my thinking on the problem this
is one that I don’t know if I agree with
but it’s a really strong statement and
it seems like it could be a pretty good
heuristic this is by and ruing who is no
longer the chief scientist at Baidu
but he says if a typical person can do a
mental task with less than one second of
thought we can probably automate it
using AI either now or in the near
future I can’t think of that many
counter arguments that don’t require
like a lot of like very specific domain
knowledge like maybe people who play
games a lot can play those games really
fast cuz they’ve practiced it well but
roughly it seems like a pretty good
heuristic and it’s a very like strong
statement so maybe this is something
that could guide you answering that
question another thing that isn’t as
specific but I think is very important
to consider is that a lot of the deep
learning successes today have been I
haven’t used the word simple but I feel
like they’re more simple memorization
problems and not really thinking
problems it’s always hard to really say
what this thinking really means because
that might be a moving goalpost of like
of course if algorithm is not thinking
it’s using an a-star algorithm or
something while we might think that’s
kind of like thinking but even in this
case it seems like when you’re when a
task requires multiple steps of
reasoning where you can’t like use
heuristics to jump all the way from
input to output it seems the deep
learning has not been very good at that
especially not without a lot of help
which leads me to my general rule which
is deep learning is an appropriate tool
for supervised direct pattern matching
tasks bonus points if you can design
priors that are particularly suited to
your problem the the priors in this case
are specific layers that are popular for
certain tasks but we don’t have to get
into that right now
even though feel free to ask me in the
Q&A but here when I say supervised I
mean that we tell the model directly
what the correct answer is so roughly a
human or some other process figures out
what the right answer is via some means
and tells the model this is what you
should be outputting next time there are
there have been incredible successes
using reinforcement learning which is
not supervised especially in the
game-playing domain so if you’ve seen
deep Minds deep queue networks playing
Atari or deep Minds alphago playing go
those use a lot of reinforcement
learning and if they definitely have in
some successes but how I feel about
reinforcement learning is that it can
work but you don’t want to rely on it
working and in the bay everyone hasn’t
started up on everything and there’s
been a lot of people who kind of have
bet their companies on this is a
reinforcement learning problem let us
sell people on using deep reinforcement
learning to get this working and ending
up with vaporware and kind of a sad end
to that supervise for it but as far as
direct pattern matching goes this goes
back to what I was saying earlier where
you want simple relationships between
the input and the output
almost almost like that a fraction of
the input directly maps to some fraction
of the output in some sort of additive
ish way it doesn’t have to be completely
additive but usually having some easy
mapping allows you to bootstrap the more
complicated mappings and a lot of the
more complicated mappings turn out to be
like lots of little simple mappings
composed together and this kind of thing
seems to be how deep learning tel tends
to work this is all very vague but I am
about to talk about some specifics about
where deep learning has succeeded and
where it seems have not succeeded yet
another disclaimer this is only a subset
of the potential cool things to talk
about and I’m only talking about the
intersection of things I find
interesting because I want the slides to
be interesting there’s lots of like
little things that are cool but maybe
wouldn’t be that interesting to people
and visual and things that I know how to
put in a presentation so I I have some
attempts at videos but they’re optional
but I there’s some cool lots of cool
working audio but I just have no idea
how to put that in a presentation and
[Music]
yeah maybe that’s my bad but we can
solve go and build robots but technology
isn’t there yet for reliable audio and
video this is also what happens when I
make the graphics myself so all of the
pretty animated graphics have been
stolen from other Googlers who know how
to do ours
back on topic let’s start with the
easiest thing whenever you have a metric
that when that metric goes up money goes
up you probably want to use machine
learning possibly deep learning but
definitely machine learning this is
actually I would describe the main use
case of machine learning
I have knobs to turn some combinations
of these knobs are better than others
how do I turn them basic stuff but we’re
saying a big thing that people get
caught on is unsupervised learning it’s
a very interesting research problem but
if you want to do anything practical I
would probably advise you to not do that
I think this is actually absolutely
excellent advice rather than trying to
if you can spend a month figuring it out
on supervised learning please do it that
that will solve a lot of people a lot of
time if you could spend a year that
would probably save a lot of people out
of time if you could do ten years you’re
probably on track with the rest of the
field so if you have a problem that you
care about don’t try to do some magic
where you don’t know if it’s gonna work
label some data usually these things are
a lot more data efficient and people say
they are and sticking to supervised
learning will be much easier for your
sanity as well as your eventual impact
speech recognition has done really well
really really well people think that
this is probably going to be one of the
biggest changes to interfaces in not
just our lifetimes but in the next
decade right now people don’t like to
talk in phones because they could be
kind of sucks but people can talk much
faster they can type and a lot of people
don’t know how to type very well so this
could completely change the way people
interact with electronics things like
Google glass or I hear it’s really big
in China speech recognition there’s all
sorts of things that this could enable
and this is only a fraction of the cool
things happening in audio but I not
think to talk about that much but
there’s things with generating
generating audio generating music lots
of cool stuff there translation this
animation is really cool and this
problem is really cool this is showing
that not only can deep networks improve
on like the traditional statistical mess
that things like Google Translate used
to do but where you just have matching
corpuses or Corp I I don’t know what the
plural is but you can also translate
between language pairs that you’ve never
you don’t even have matching corpuses on
so in this example you have English to
Japanese pairs as well as Japanese so
sorry English to Japanese in English to
Korean and using these networks you can
actually translate directly between
Korean and Japanese without ever seeing
paired data between Korean and Japanese
which is actually huge
it could enable a lot of translation on
languages that between languages that
there’s just no data on and you can do
it in a much more accurate way because
you don’t need to translate into an
intermediate language where you lose
some information if you ever like do
Google like what’s the game where you
have a like a Markov chain with Google
Translate you start with a thing you
translate to one language you translate
back eventually becomes garbage and
nothing like the original thing you said
and you just avoid that problem entirely
with this image classification this is
like the bread and butter of deep
learning it’s what made deep learning a
big deal people it was kind of a not
mainstream thing until about two twenty
twelve when deep learning one this
imagenet competition and beat all of the
other things by a fairly large margin
and made everyone realize hey this
solves problems that nothing else could
solve before and there’s real-world
applications to this like google photos
an example there’s a lot of api’s where
people have made a business of telling
you what’s in an image people do face
classification Faith’s
detection there’s a lot of money and
sentiment recognition you know like have
a camera here and look at the room tell
them if they’re enjoying the talk or not
based on like people’s smiles and stuff
maybe not for talks but like for ads and
stuff something that can’t be done yet
though is unbiased image classification
or it’s still a lot of work this was a
huge issue for Google photos actually
like I think it’s like a few days after
they released it people were complaining
on Twitter that
their friends were being classified as
gorillas due to a lack of diversity in
the training data and this is kind of
unavoidable when you have imperfect
datasets I actually don’t know how they
solve this they might have just removed
some of the classes that could be been
taken as offensive but that’s just a
hack right like we want like real
algorithms that don’t make these kinds
of stupid mistakes talking about not
making stupid mistakes a problem near
and dear to my heart is medical imaging
they’ve been a bunch of huge successes
on medical imaging in particular there’s
been some really cool stuff done reading
x-rays and CT scans cool stuff with
segmenting pathology scans detecting
diabetic retinopathy all of these things
it’s people have been getting superhuman
results like better than what seems to
be the best doctors and hopefully very
soon this kind of stuff will be like
reaching the end users and helping
people so this is a really exciting area
of deep learning progress similar in
that vein it’s not limited to either 2d
images or having a single prediction per
image you can do what’s called semantic
segmentation where you label each pixel
or in this case voxel in an image and
you can also it also works for high
dimensional data so this for example is
3d segmentation of I believe in neuron
and this algorithm actually is iterative
and how it like expands over time and
this is very similar to how a human
would segment a neuron it would not just
say all those months here’s a neuron if
it starts at something and being like
okay this is close to this other thing
this is maybe a neuron so we are as we
were like expanding the reach of deep
learning more people are designing more
and more of these priors to build into
the architectures to do much smarter
things so whenever I say not yet on
something it might be that technology is
there P we just haven’t tried hard
enough talking about and not yet there’s
been some really cool work on image
captioning so instead of given an image
output a object in the image it’s given
an image describe the image
and this is a much harder task because
there’s a lot of things that can go on
in an image and there’s a lot of
possible ways to describe an image so
how do you say something is riot and
what what set of things do you choose to
have something described and this is
pretty good like these descriptions are
actually this is a good case there’s
many bad cases of this but they still do
make some really dumb mistakes it might
reflect underlying issues with our
imaging models or it might be due to
dataset size but this is still an open
research problem similarly to that it’s
not very good at answering questions
about images or stories it can be good
at finding specific things in the images
but there’s other things that seem to be
easier than finding a thing or just as
easy as finding a thing that deepening
currently it’s not good at like counting
if you ask good this is I don’t have a
counting example here but if you have
like a bowl of oranges and you ask like
how many oranges are in this bowl this
sounds like a very easy task but it’s
quite hard for models right now so
that’s a big problem talking about big
problems we definitely are nowhere close
to automating research this is a great
tweet we’re the researchers were the
ones that wanted to make the AI do all
the work and play games and while they
play games but instead it’s the opposite
right the yeah is just playing games all
day and researchers are working harder
than ever it’s a tough life I think the
comments on this were equally great
because maybe this is a sign that the AI
is actually intelligent you know maybe
it’s like just pretending to be dumb and
being like why would I want to do all
the work I’m just gonna keep playing
games all day some aspects of research
might be automated something that some
people consider to be either boring or a
waste of time or hard is designing these
architectures in the first place and
there has been some work in using deep
learning to automate the architect the
design of architectures for more deep
learning and you get like these crazy
things that no one would ever design
yeah I would definitely not
think to do that in the right so this
stuff has had some fairly promising
results I I put this under a maybe of
what can be plausible because it’s both
it was both very expensive and not quite
as good as a state of the art but this
seems like a really promising Avenue and
a potential place that it could make a
big impact so maybe all of our learning
about architecture and studying this and
trial and error maybe all of this will
be outsourced to you know farms of
computers somewhere and we could just
you know stick to the high level tasks
but life is rarely that kind a despite
fake news to the contrary we are long
ways away from automating software
development there were some articles on
algorithms automating coding and III
think that some people were a little bit
panicked on this maybe all of the
articles when deepening automates X
causes some panic but I hang around I
hang out with lots of software engineers
so they were worried for like a second
until they realized that this thing was
actually really really dumb not that the
the work was done but how the algorithm
did it was nowhere close to software
engineering it was a slightly better
heuristic for picking random bits of
code together and doing trial and error
on that code and as we all know that is
absolutely not how we do software
engineering right like we design stuff
upfront not trial and error it’s like
all done by the books yeah this
algorithm definitely can’t do that
so we our jobs are safe right guys
ok so some people some people know what
I’m talking about some great memes this
is not really model output but if you do
follow the field people love to have fun
things in there some people actually
wrote a conf a fait paper I think that’s
pretty incredible it’s someone can like
dedicate a research paper with I assume
a real idea I didn’t read this but I
assume it’s a real idea
to a troll name I think it’s great and
it shows like the speed of publishing in
the field common Silicon Valley problem
no deep learning will not solve all of
your problems especially not your
product definition problems it won’t
find something useful for you to do and
it will not make you magically rich
despite a lot of belief to the contrary
and similar to this image general chat
bots are actually quite difficult you’d
think that you just give you no model a
dataset of two people talking and it’ll
be able to replicate those people
talking but it turns out that our
language models are quite good at making
things that look grammatically correct
but are semantically quite terrible so
they don’t have like they don’t have any
history involved they don’t have there’s
lots of issues of them and this like a
misunderstanding too that led to a lot
of companies starting products that
ended up pivoting away from using deep
learning at all and ended up using like
an army of workers in the Philippines
this manually doing the chatting for
them which turns out to be a pretty
economical way to do things but specific
chat bots are very doable so when if you
turn the problem from hey let’s generate
arbitrary text to hey let’s pick among a
small set of valid responses things
become a lot easier this is in boxes
smart reply which apparently is used by
over ten percent of mobile infox replies
which sounds like a lot of
qualifications but I just think it’s
cool that something started out as an
April Fool’s Day is now real April
Fool’s Day joke it’s also sweet
animation but this kind of stuff is very
plausible and I think that people who do
use machine learning for chat BOTS will
end up constraining the problem quite a
bit and that’s actually very doable if
you’re trying to classify like do I have
enough information or is this person
satisfied or do I need to pull another
human in to actually chat through this
person that’s much more doable than hey
automatically solve this person’s IT
issues which sounds really hard and
similar in vain to that coherent text is
quite a challenge like long any long
amount of text a lot of journalists
journalists I feel like a big victim to
hype because it’s kind of their fault
uh-huh and they’re kind of worried about
their jobs about RDR deep nets and like
start writing articles for us and the
answer seems to be no so if you’re a
journalist
don’t worry sorry
did you say it’s hard to investigate
journalists as an AI
I couldn’t quite hear the last part of
that oh yeah for sure
he said as may I it’s hard to do
investigative journalism that is
definitely true debatable about how much
investigative journalism current
journalists do but that’s definitely the
case I think in this case it is even the
worry that a lot of journalism is read
stuff on Twitter turn it into an article
hope to get lots of clicks make click
vadie headline topic definitely not all
of it but some fraction of it is that
and I think that there is some worry
about this like I believe in finance
there’s a big race to like who can
publish these articles first based on
various data sources and if you’re not
talking about quality but speed these
things definitely have a speed advantage
yeah III I’m not I’m personally not
worried about journalists jobs being
taken for sure cool sorry sorry I
couldn’t hear you very well but thank
you for yelling that time yeah so
something that seems to be really
promising I actually think that this is
one of the most promising upcoming uses
of deep learning that’s like not quite
there but might be there and it’s like a
would be a really sexy field to get into
with robotics it seems like there’s a
lot of really good stuff happening with
imitation learning and a lot of people
are invested just working oh it is
working this is pretty cool people are
investing in a lot of the the research
labs are investing and getting the
robots training together in like how do
we collect lots of data for robots to
get them working automatically because
right now
at least to my understanding I’m no
roboticist is that majority of the work
done by robots is done manually and if
we can like make it a lot easier to
train robots to do things that we care
about maybe
all of a sudden we’re going to have like
more general programmable robots that
people can do stuff with so that I think
is really really promising and at least
from a research perspective of someone
who reads the papers and like keeps up
with what people are doing it seems very
plausible that this kind of thing could
make a breakthrough in the near term
especially with what’s called imitation
learning where robots rather than
learning by trial and error which can be
very hard they just learn to copy humans
which is goes back into the rule of
thumb I was talking about we’re giving
giving these algorithms supervised data
telling them what to do generally works
a lot better than hoping for magic that
you know hoping that they will magically
figure out the thing to do which is what
a lot of the field is trying to get
working right now depending on who you
ask
game playing I would count categories
that is not yet there there have been
some amazing successes in game playing
but a lot of those successes aren’t
quite super general a lot of it it’s
like very input simple input output
mapping like I was mentioning so Atari
seemed to be a lot of that debatable
whether or not go was that even though
that was definitely a huge win but
there’s been other games where models
are nowhere near as successful so things
like even like very simple Minecraft
mazes it’s still the model still aren’t
quite there yet or recently there’s been
a bunch of work on doom visual doom like
do them from the pixels and this model
actually super cool let me see here does
this work you can skip to the fighting
so there’s been a lot of progress on
that really recently this is like the
what was the state of the art in 2013 if
you can tell it’s like pretty dumb like
shooting a wall right now let’s see here
yeah this is so this is a little bit
smarter this was state of the art in I
would say 2016 ish mid 2016 this is
still pretty dumb
and people been making a lot more
progress recently with this we’re look
at this this is actually intimidating
it’s like moving around its shooting
intelligently etc etc there’s been a lot
more progress being made in this and it
seems that we’re nowhere near close to
or at least to my knowledge solving
something like Starcraft but it’s really
promising and people are putting a lot
of effort into this so the likelihood
that we make some big breakthroughs in
the coming years seems to be likely cool
and this is category of stuff it’s like
what I am one of the most excited about
just because I would never consider
using like these extremely powerful
classification models for artsy things
maybe that’s just me but some of these
use cases are incredibly creative and
incredibly cool and like this stuff is
amazing this one came out pretty
recently and it learns to transform
images from different domains so
transforming like a zebra into a horse
or vice versa so image transformation it
can be done can even be done with videos
so this is like actually really done by
a model it’s not like cherry pick data
but so it’s actually a transform this
video and like this is not seamless but
that’s pretty good better than I could
do with Photoshop which is not saying
much but like this is like pretty
impressive and I would not even have
thought of this as a use case like hey
I’m a machining researcher at Google I
have a you know giant cluster
I’m gonna transform a horse into a zebra
right
unfortunately this kind of thing is not
completely reliable but this is pretty
amusing so like with all machine
learning like making it completely
reliable can be challenging talking
about that there’s been an app that’s
been gaining popularity called face app
that
does facial transformation so in the top
left you see the original photo top
right you see like a more manly
transformation you know like more edgy
chin bill beardy then bottom left and
old miss transformation and bottom right
a smiling transformation and this is
actually pretty good pretty good and
like an app can do it on your phone no
human input it just does it it’s pretty
impressive that you can do this and this
is a really cool use case unfortunately
it’s not perfect in particular it’s also
suffers from that bias problem like with
many other things when you turn a cool
model into a product it there’s a
different set of requirements in this
case they had a transformation which
makes a person’s face hotter and one of
the things it did was it always lighten
skin which was offensive to some people
yeah that they had to pull that feature
I think or change I think they actually
changed the name from hot to something
else I can’t remember art is doable this
is art from scratch or unconditional art
like it’s like these models can just
create these artsy things and I think
that this is really really cool I
personally think that these both look
really good can I get a show of hands of
who thinks the one on the left is better
what about the right oh it looks like a
tie I made the one on the left so I was
hoping that people would vote for that
one cool I think they’re both really
cool I would definitely have a poster of
that in my room or a painting of that in
my house then this kind of thing like
who would have even thought that as a
side effect of these really powerful
actually useful things we would get art
oh yeah I’m definitely not claiming that
this is the like the the the first thing
in terms of algorithmic arts it’s just I
find to just be a really cool use case
of deep learning because when III don’t
think ten years ago people would have
imagined like yeah imagine all of these
cool pictures will make and every
actually every time there’s a new use
case in art I’m just amazed like who
thought of this like who spent their
time on this and I’m thankful for that
because I wouldn’t have done it but I
think it’s really awesome and I think in
some ways it’s also kind of cool that
unlike fractals or something like that
it feels like there’s there’s a lot more
like there’s more unknown unknowns in
this right now which makes it really
promising as well maybe that’s from my
misunderstanding of art though or
algorithms or anything I’m not expert in
any of this stuff sketching with another
recent use case where you just train in
a data set of humans drawing little
things and the things in the the top
corner up there is things that the model
drew and in the bottom here you can
actually do math on sketches so you take
like a cat face you add in a pig with a
body you subtract a pig face and you end
up with like a cat with a body and like
it’s kind of cool that it works I mean
the math checks out so awesome for for
the non artists in the room you can also
turn what is arguably not art in the
bottom-left into something that is
potentially art so this is another very
cool use case where you like it can like
enable people to it becomes almost like
you know a new artistic medium right
where you can now use these things to
enhance existing art to do things maybe
that people wouldn’t have done before or
enable people who couldn’t have done
this before
or maybe just make it more faster
thing like that it feels almost like you
know like a new instrument from the
musical sense so this stuff is really
cool style transfer I think this is
crazy because a year and a half ago this
was already looking really good and it’s
just gotten better and better so this is
like going so well I actually I should
have put the old pictures here as well
but like these are the new what I think
is the latest in style transfer and this
is pretty good like you can see
transferring the style of a fire into a
bottle like this is like a professional
Photoshop job and I guess and this is
like impressive and I would want to do
this and I look forward to this being
able to be done for me because I don’t
want to implement it myself but there’s
a lot of really cool stuff being done
with style transfer and this stuff is
really pragmatic because like
aesthetically this is already like very
high quality this is my crowning
achievement actually mixing my face with
that of a Pokemon probably my best
achievement and deep learning definitely
works would recommend trying it again
and probably newer stuff will work even
better and as far as specific go there’s
all sorts of other things a rough
formula is taken input that is similar
to another input that deep learning has
succeeded on like images audio raw text
other domains like that pick a response
that is a relatively simple mapping from
that input so nothing too complicated
but simple mappings like are they human
faces and collected data set trainer
model usually something like that gets
just something that works quite well and
as far as what it can do if you if you
pick the right things generally it
generally makes it is the valgar isn’t
helped you a lot in doing
a lot of the easy work for you getting
like the last bit of presents always a
lot of work but you you’ll know if you
can get it which is it which makes it a
little bit easier
oh so back to the big questions how will
the world change I think this is a great
tweet like Andrew Inge
I do believe automation and steroids is
the right way to think about it not
sentience or AI overlords or anything
else like that I actually would be quite
pleasantly surprised to see generally I
in my lifetime just because I think
that’s so unlikely and that’s not
because despite what my current pants
might imply that I’m one of the live
fast die young types it’s it’s that I I
think that it’s quite quite a ways away
though I would love to be wrong as far
as how the world change I won’t claim to
be an expert on the societal effects of
automation but luckily this guy would he
had a TED talk called will automation
take away all our jobs all seems like a
little bit of a weasel word here it has
over a million views and bet riches law
of headlines applies here any headline
that ends in a question mark can be
answered by no so the answer is no
saving you 18 minutes the claim is that
AI automation just like other automation
will take some jobs away but it will
probably do much more transformation of
jobs because a lot of jobs aren’t these
just simple mappings and there’s more
complicated new instincts to them but
automation increases the leverage of any
person and there will be a lot more jobs
that we just can’t imagine will happen
so that’s his view I don’t have strong
views in this yeah what can deep
learning do the fields really exciting
there’s a whole lot of things that we
can do now that we previously couldn’t a
lot of things that were once thought to
be really really hard are now doable
people fought out of solving go was like
a hundred years away or more and there’s
all sorts of fields that this could
affect
as far as specifics ago which is what
would actually be more useful to you
guys
the answer is it’s kind of complicated
my advice would be to look at example
failure and success cases network with
researchers and people in industry or
use someone’s rule of thumb maybe my own
maybe take it and change it but really
you want to build your own mental
classifier of what isn’t isn’t possible
and refine that classifier around the
set of problems that you specifically
care about so if you want like a
specific like I want to figure out if
deep learning can find X in a genomic
data set you like go into the research
look at that figure out like what seems
possible what’s not and there’s so many
problems out there that you might have
you might end up being the world expert
in knowing if deep learning works for
your problem so it’s just there’s so
much opportunity right there that if you
ask anyone there probably will tell you
an answer because people love to give
answers but they probably won’t give you
a very good one including myself so
right now it’s unavoidable to do some of
that work unless you do something that
someone’s already solved but that’s kind
of a cop-out answer the last big
question it was how do I take advantage
of these trends I think it’s a lot like
learning softer engineering especially
back in like when the internet was young
and my answers scratch your own itch
play around with it a lot of the work on
art specifically was done by hobbyists
and not researchers and we have no idea
yet what can be done on you know the
problems you care about and for all you
know you might be sitting on depending
the next killer app that no one else is
thought of and it’s scratching your own
itch leads to something valid for others
start a company on it there’s lots of
money for companies going around right
right now and the world needs more AI
companies that actually provide value
I’m not gonna name names and also
prepare for like a sweet transition
consider joining Google it’s the best or
any other like AI focused company which
is like interesting impactful problems
and the resources to solve those
problems
thank you
you
Please follow and like us:

Be First to Comment

Leave a Reply