Machine Learning Research & Interpreting Neural Networks Video Lecture | Coffee with a Googler - IT & Software

[Music]
so Christopher welcome to coffee with a
Googler it's great to have you on it
thank you for having me
so I know you are one of the people you
work on Google brain and you're behind
this distill for a great dissemination
of machine learning could you tell us
all about it
yeah so I guess maybe I exploited a sort
of into into two parts okay so they're
sort of the thing that that we're doing
and then there's sort of how it is still
the vehicle for that so we have this
idea of there being technical debt right
you can write software and you you maybe
you don't always use good variable names
you don't document it very well you
don't you know you end up with all of
these you know you're you're so you
don't refactor it and your your software
becomes hard for people to go and build
on and work but I actually think a kind
of analogous thing can happen in
research where if we don't explain
things well we don't sort of develop the
right ways for thinking about things it
can become harder to go and build on
research and it sort of becomes this you
know this mountain of papers where you
know people and I think people take
pride in it right they're like you know
I spent all these years going and
climbing this mountain and I got to the
top and I'm gonna do more research and
build up this mountain um but I think
often we actually could have done a much
better job of making the mountain sort
of easy to climb yeah
and so what still is really trying to do
is to create a vehicle where people can
do that kind of work and and do really
amazing work of that kind and have it be
acknowledged as a real sort of
contribution to the community and and be
the sort of thing that they can then
help them get supported in in doing more
of that work so it becomes like a really
a best of both worlds kind of thing
that's very accessible to the
non-academic community but it's also
referential if that's the word you know
something can be referenced by the
academic community you get the best of
both many of these things that were sort
of already just a better version of
scientific papers um but they weren't
being counted as such and so I think
it's still sort of a legit amazing
vehicle that allows and I think we also
provide some help to people in doing
good work in this in this way so there's
all this cetera weird you know
bureaucratic stuff and we just sort of
you know crossed all our T's dotted all
our eyes and and sort of set up you know
distills also peer-reviewed and and
created this you know this a scientific
journal except it's a scientific journal
full of interactive diagrams where you
know we really you know really if
somebody's writing writing in a very
academic style we might you know
urge them to try to do a little bit less
of that right hum and and sort of
subvert what a paper isn't in some sense
and sort of and I think sort of explore
what what's possible right I mean I'll
talk about from my personal perspective
and that is that like you know I have
been out of academia for a long time and
I've been working in software and
software development that kind of stuff
but as I've been getting into machine
learning of course I have to start
reading more and more papers and as I'm
reading those papers like as soon as I
start getting into mathematical notation
and sigmaz and fives and that kind of
stuff
part of my brain begins to go to sleep
and it's not that I don't understand it
it's just it's so alien to me it's so
long in the past and you use the image
of a mountain and it's like I used to be
kind of three-quarters of the way up the
mountain now I'm just in the foothills
and it's a lot of effort to get climbing
again so if it's something like distill
and what you've been doing what the
still has been great for me personally
so that I can start really beginning to
grasp some of the concepts and and
speaking those concepts is like another
thing you've been working on is lucid
right oh so and you're using distill to
disseminate lucid could you tell us a
little bit about that or yeah so I think
why don't this really really traces back
to to deep dream not even a little bit
further you know there's there there's a
there's a lot of people but there's a
small cluster of us who've been working
together who've been very interested for
a number of years and you know how do
neural networks really work and how many
humans understand these sort of really
really complicated systems um and you
you know a couple years ago we had these
really exciting results where we we came
up with all these techniques for sort of
visualizing you know what what do neural
networks think are I think are
interesting what are they really saying
what work I mean really really think the
right way to think about those images is
they're sort of they're you know they're
trying to make the most interesting
image for our network as possible and so
it gets filled with all the sorts of
things the network you know finds finds
very very interesting in some sense that
writing that activates it all right now
it's like it's full of you know dog
slugs and that's a lot of fun and it has
even become this this artistic movement
we there was this event at the gray area
foundation where people were like
auctioning off you know art that was
made with deep dream and I was like whoa
you know that's right this thing that I
was involved in is like there's like
real artists so we're doing things I
hope you have some of those
I think we're all too expensive for
anybody so deep cream you know I think
part of what's exciting about it is is
all this art um and you know I think we
were we were really excited to are
really lucky to you know see there's all
these amazing videos but you know in
addition to that I think you know we
were you know it was also sort of really
an attempt to understand what's going on
inside neural nuts right and it didn't
did stop there um you know we we sort of
continued thinking about this a lot and
then a you know about a year ago or like
a couple months ago and we we published
this article on future visualization
which is sort of really exploring how
can we take these techniques that we
started developing with deep dream um
and then go and turn them into you know
turn them into tools that are sort of
instead of just trying to make
interesting images really understand
what what individual neurons in the
network are looking for okay
and we discovered you know in you know
early layers it's looking for edges oh
come and then you know it starts to look
for textures and then it uses the
textures to describe you know simple
patterns like stripes and fluffy balls
and logos and hexagons and you know
clusters of circles and you know these I
don't even like me like spear like
things and then those get turned in to
you know simple sub parts of objects
like you know a button detector or a
flower detector or a patches of cloth
detect or a bubble detector or a chain
detector or a nose detector and then
those get turned in to sort of partial
you know part you know partial objects
like you know parts of buildings or
people walking around or helmets or dogs
with fluffy or floppy ears and little
insect like things right and so that was
sort of a really exciting transition and
then it sort of was the the result of us
started fiddling a lot with ideas around
keep dreamin not just ASEAN there's been
a there's a big community people have
been doing really exciting work in this
area and we sort of have been been
building on that and building up
building up infrastructure really you
know that just is just a means to an end
where what what that work was giving us
was it was giving us a way to go and
understand what our individual neurons
doing okay individual neurons in the
networks those individual neurons that
seem to know detect floppy ears and
things like this
okay but then what we've been doing
lately um is we've been going and taking
about a step further and going and
saying you know how can we then use that
as a building block in conjunction with
other things and to try to really
explore you know how the network makes
decisions so I start with a simple
example um
so over here we we have an image right
now we're looking at an image of a
fluffy you know this Labrador Retriever
and this tiger pad okay that's actually
apparently a species of cat so there's
been called a break okay yeah I mean
there's real Tigers and then there's
also tiger cats which I guess maybe they
have stripes on them or something but
this also you know like you have to have
you had to have tiger cats apparently
that's a tiger captain I would have
known what the network does but the
network does it's really amazing I would
have actually wanted knowed it was a
Labrador Retriever either I know the
network really it's it's you know it's
it's pretty phenomenal you classify as
you know dogs into more than 100 species
and you know I said I don't even know
the names of a hundred species of dogs
let alone you know how'd it go and tell
them apart but the network can and one
thing you might wonder is how does the
network do that right um
and one really exciting thing that we'll
see is well let's just dive into this
for a second sure if you want this this
interface here it allows us to to go and
look at the vectors and then the neural
that runs the theme detectors at every
position in the image um and normally
that you know those detectors they all
given a number at what the sort of
describes how much they fired and
normally you get this so called
activation vector and it's just a list
of numbers and it's sort of really
inscrutable you're like you know neuron
53 was firing a lot well what does that
mean the buffer useful to me doesn't
tell me very much
you know neuron 134 are also fired about
a bunch but neuron in our own new neuron
12 it didn't fire well great thanks but
that doesn't tell me anything there are
on 12ax7 about yeah about what we can do
is we can combine it with the these
feature visualizations okay that give us
a sense of what what the neurons are
looking for mm-hm and we can combine it
we can create this we call them semantic
dictionaries where we have you know
these these things that sort of give us
an a name for the neuron may give a sort
of a visual symbol that describes what
they're looking for with how much they
fired and that's that's a lot more
informative so if we look over here you
know all over here we see the disney's
floppy eared detector
that are are firing really intensely um
it turns out that you know this neural
net and Cokely net has a really rich
vocabulary of different kinds of year so
that's a really big part of how it tells
the part animals so
different kinds of floppy ears you know
it not just floppy ears but you know
this these longer floppy ears it's these
mid length floppy ears and it has all
sorts of pointy ears as well and like
you know slightly it's like yours that
are sort of in between pointy and fluffy
um here what we're seeing that these
really these two floppy ear detectors
that are firing pretty strongly um
that's really interesting but it's
something that would have been
completely opaque to me right without
going and combining these techniques
together right um and so that's sort of
the first thing we're doing we can look
in other places so there's a snow
detector over here and that really is
firing pretty strongly and if we go over
here you know man we're seeing some
point to your detectors and sort of a
cat face like detector and you know
that's what almost looks a little bit
like an ape detector that's firing a
little bit a little bit over here
actually go down here we see you know
this these fur detectors and these leg
detectors that are firing you go down
here to the grass you notice some grass
detectors but they don't fire very
strong likes that's not very interesting
and we can look at other images so you
know over here we have a fellow with a
with a you know with sunglasses and a
and a bowtie and you know turns out
there's a bow tie detector oh and
there's also you know a net without a
bow tie detector and what I really liked
about this one was that like the bowtie
and sunglasses look similar right it's
too dark oblongs oh they smell
brilliantly and but this was actually
able to detect the whole ya know it's
really it you know you can see here and
you can see that part of what part of
what it's doing is it's like you know
it's looking for the chin above the the
the bowtie um and I think also for sort
of the shirt with buttons a little bit
below it now so these other things that
are cueing it but it it actually is a
pretty you know sophisticated sense of
what a bow tie is and all says the suit
detector and yeah the the neck detector
and yeah here's you know some kind of
face detector that seems to mostly be
about like noses and mouths and skin
texture and you notice the sunglasses
detector up here and no let's look at
this one and it turns out that this has
you know these you know top of these
detect our body of veins detector
there's sort of a flower detector
there's an
detector that helps it detect the handle
and interesting yellow spheres and that
was another thing that was interesting
in this one is because the lemons at the
bottom of the image are yellow oblongs
and the tulips at the top are yellow ha
ha ha
you know despite those two being very
similar that it's able to detect that
there are different things yeah and in
fact now if we well I'll go to that in a
second but there's there's some other
one there's some we can see some
interesting things about how these all
sort of play together in a second but
yeah so I think the first thing is just
you know this is allowing us to
understand you know sort of looking look
at which neurons are firing and and see
what's going on I'm gonna show you the
pink for a second the pig is a great
example of something that has lots of
pointy red actors move around a little
bit to find a place where that we're
really getting all the point ear
detectors to fire this too but now in
this some places were more fire well
maybe I'm not gonna be patient oh yeah
there's a bunny here okay look the wave
point to your point to your point ear
and then like some kind of rabbit like
thing with me with another kind of point
you it has this whole vocabulary for
talking about you know different kinds
of your anything like that right so
that's that's the first step um we we
sorta we wanted to go and take that a
little bit further so another thing that
you can do is you can ask you know watch
all these neurons together represent and
so we you know there's all these neurons
that are firing a different extents and
if we instead of just trying to
visualize individual ones we ask try to
sort of visualize them all together we
sort of see what the network sort of
collectively saw at that position and
then if you stitch those together you
can kind of see how the network saw the
whole image so you know here's our you
know pig or hog or whatever it is or our
cat and our dog and you can see you know
what really seemed that the ear here and
the snout and the the cat head and the
legs and the grass or or over here you
know the the top of the vols and the
handle and the flowers and the lemons
down here and the wooden surface in the
background and or or over here you know
there's the bird on the chain and it
really sees the bird's beak and the
feathers and you know the chain and all
of us I mean that's kind of interesting
pretty incredible and then we can
actually go a step further and the way
that these networks work remember is
they they multiple layers they build up
their understanding over the course of a
bunch of layers
so initially there really you know
what's zoom in here you know initially
they're really focused on relatively
simple things in fact we don't train you
the first layer but it would really be
about edges right and you hear it sort
of you know it's getting really simple
you know combinations of edges or maybe
the beginnings of textures but we go
down no just two more layers we now have
we're skipping a layer in the middle
here we know how you know this thing or
it's really got a much richer sense of
textures and it starting to get a bit of
3d structure and if you go a little bit
further down now we're seeing the chi
and the beak and so on and if we go a
little bit further it's sort of become
more abstract and sort of birds and
chains now similarly if we go you know
let's look at the dog I had left the
stalk in this cat so you know here again
you know very simple sort of patterns
with really mostly about edges but you
know you go just a step up and now
there's fur textures there's grass
textures there's you know this you can
really see that it's this 3d structure
where there's boundaries and sort of um
you know surfaces a little bit um and
then you know another Stefan it's
starting to understand eyes too
um but now you know it's really got like
the snout with eyes the ears the the leg
and so you can really see any job that
sophistication of understanding the
object as you go up and this is part of
the magic of distilled right so this
isn't just a dry academic paper with
illustrations in it we can see you
interacting with it we can see you
actually seeing what a neuron is seeing
and getting hands on to that so it
helped me to learn a lot more about
what's going on something extremely
powerful allowing people to you know
there's sort of these pictures that I
have in my head and there's I think it's
sort of actually two interesting things
going on so one of does these pictures I
have in my head
right I really turning them into
interfaces um I'm sorry I go able to
crystallize those pictures into things
that I can really test my intuitions
against right and that kind of um you
know offload some of the thinking that I
have in imagining what this would be
sort of what's going on in the network
and really rheya fiying it really
turning it into this thing that I can
interact with mm-hm and then normally do
I keep that for myself
but I can share it with other people and
so they get to now see you know this
thing that's I think deeper than what
one would normally get work it's not
just me telling you results but I'm
really starring this way of thinking
about and sort of interacting with with
these this this type of problem right um
and I think that's that's really really
exciting I mean cuz to me there's
there's two levels of opaqueness with
machine learning right first of all it's
learning it to begin with you know which
is already opaque and then secondly is
like when you start dealing with models
it's it's a black box right it's a
there's a there's a trained graph in
this thing that you give it a picture
and it tells you what's in the picture
so there's a it's not like source code
that you can open up and step through
and see what's going on so though
between those two levels of opaqueness
it's really really hard for somebody to
get into this and to learn and what I
like about this is that this is cracking
open the second of those I can begin to
really see what's happening in the
network and help me understand the
network and then later on help me to
build and tune my own networks all right
so which is which is super cool so now
lucid where does lucid fit into this
picture so with all of this you know
even we've been building up all you know
we building up all these tools and we've
been building a lot of infrastructure to
go along with it and so one of the
things we're really excited about with
this paper is we're also open sourcing
all of all of the infrastructure that
we've built up to go and do this
research ok um and so that includes both
that that first to feature visualization
article that I was showing you where you
know all of the the tools you need to go
in and produce images like this where
you you visualize what a neuron is
looking for and then also to go and you
know that's also what underlies all of
all these tools that we have here and
the neurons in this one are they just
trained on something like inception or
yes these are for both articles we're
using this network googly net okay which
was at one point to stick to the air
model it's now separately several years
out of date but it's one that we sort of
have used for a while as all sort of
standard test for for visualization okay
um seems like it actually there's some
kind of something a little bit
mysterious matter where it seems like
the neurons in googly nod are like
especially correspond to ideas that sort
of human sort of our meaningful to
humans I say Tom it's a fun one to start
playing around with cool go home but you
know you can go and plug in whatever
model you want
um and try and use these techniques on
it nice nice
she made these notebooks that reproduce
each one of the the diagrams in the
overhaul so you can go get hands-on yeah
so you can you can just sort of you know
it's there's sort of this continuum
where you can even just read the paper
passively you can you can engage with
the diagrams and you can go a step
deeper and go and start playing around
with one of these notebooks and you know
so here we have you know a notebook for
these these activation grid
visualizations and then you can go and
and you know just you know I open in a
playground but it used to be like once
upon a time that you'd read a paper and
you're trying to understand what was in
the paper and you try to figure it out
and maybe they'll be a bit of source
code and you take that source code but
then you'd have to go and find a data
set of images to train but you'd have a
different data set than the people in
the paper had and they're all these like
these concepts these frictions these
little bumps that you have to get over
and right and now we can sort of make it
this continuous transition where you can
be reading the paper you like I I want
to play around with this diagram a
little bit and sort of start doing that
you know I want to go a little step
further and start actually playing
around with the code and fiddling with
it and if you want to really dive in you
know we have all of our codes open
source and you can learn the library a
little bit by playing around with
notebooks and then go and super cool now
I need you mention if you want to start
playing with all the code is open source
so say I'm a developer and I want to
learn this stuff and I want to get
started now where would I go go to
tensorflow slash lucid on github um
that's our repository um and from there
you can get access to you know tutorials
on you thinking and a list of notebooks
to go and play around with okay um and
then from you know the next step after
that is you could you know you well you
might come up with ideas after you go
and play with some notebooks and you
might want to just use code that's very
similar to what's in the notebook so you
could start writing your own own fun
things and um you know a lot of this
could also be used for artistic purposes
where you you can certainly do you know
traditional deep dream style stuff and
also your papers under still I'll put
links to them in the description of this
video so that people can go they know
yeah I think that probably what I would
do first is I would read the papers and
then I you know from the papers you can
go and and drum jump to the the
notebooks right um or or after you've
read them you can also look at the Lucid
repository and look at more tutorials
and stuff like that to play around
sounds good thank you so much
Christopher this has been
so John you know it's like we've just
we've been sitting here for a little
while geeking out about this stuff and
it's one of the things I love to do so
thank you so much and thank you
everybody for watching this episode of
coffee with a Googler if you have any
questions for me or you've any questions
for Christopher please leave them in the
comments below we'll also have links in
the description below to everything that
we spoke about today so thank you so
much again and don't forget to hit that
subscribe button and remember we'll have
a tensorflow channel on YouTube so go
check it out thank you so much