Probabilistic Machine Learning in TensorFlow Video Lecture | Coffee with a Googler - IT & Software

[Music]
so Josh welcome to coffee with a Googler
thank you thanks for having me
now you work on something called the
distributions API and that brings about
probabilistic programming what's that
all about
yeah so this is something that myself my
team were really excited about so I
think a little history is in order so
tensorflow
did a number of amazing things maybe one
of the most amazing is how they made
training deep neural nets easy
relatively easy yeah and not easier than
it used that's a lot easier than it used
to be and this has resulted in a flurry
of research and exciting results and you
know solving problems that otherwise
seemed impossible or at least very
difficult to solve right so before
neural nets were all the hotness many
machine learning folks including myself
sort of consider ourselves probabilistic
machine learning as like a
specialization and we want to be a part
of the new hotness who doesn't who
doesn't I got the logo so that's where
distributions enters the story we think
that tensorflow
doesn't have to be just a language for
fitting neural nets it can be a language
for statisticians it can be a language
for doing probability first machine
learning for fusing neural nets with
sort of older ideas and getting
something that's greater than the sum of
its parts
so distributions is a tiny Lego piece
that tries to make that just a little
bit easier now I noticed you recently
published a paper on the distributions
API explaining what it's all about could
you could you give us the brief summary
yes so we wrote just a white paper it's
sort of showing all of the cool things
that can be done when you mix
tensorflow intrinsics and this
distributions API it's actually pretty
simple but what I think the real power
comes from the breadth of the API in
terms of the number of distributions
that are included as well as the way
they sort of plug and play with
tensorflow workflows cool caller now
I'll be honest with you when I first
started seeing something called a
distributions API I was like hey cool
something I can use to distribute my
models but
yeah maybe we should have renamed it
shall we start refactoring yeah that's
uh if you want to refactor by all me
it's gonna be a bit of work yeah but so
but it's not about distributing apps
right it's not about distributing apps
no it's about encoding properties of
random variables in easy to use in an
easy-to-use framework so a random
variable is a way to represent
uncertainty let's say and there are
several useful properties that it has
you may ask what's the average under
this random variable M you may ask
what's the variance under this random
variable so a distribution just collects
these properties that's not super
exciting
I think what's slightly more exciting is
the way we organize the library
automatically takes advantage of modern
hardware okay so we have something we
call batch semantics which is a tensor
flow idea and it means that you can
parameterize one distribution in many
different ways and take advantage of Sim
D instructions right so I noticed in
your paper you spoke a little bit about
distributions we also mentioned these
things called by jack doors right that's
a fabulous name but what exactly is it
by factor so we actually named it by
Jack there so that way if you google
searched for it you would only find that
word a by Jack there is a very simple
idea it is a function of a tensor
alright and where this fits into this
you should api is presumably that tensor
is itself a sample from some
distribution so a by jekt or is a way of
transforming one distribution into
another distribution okay that's useful
because if we didn't have that concept
we have like 60 distributions now we may
have 600 distributions if we tried to
implement all transformations of all
distributions right there's something
called a log normal which is e to the
normal distribution let's say okay and
we don't actually implement that as a
first-class citizen to achieve that you
would transform a Gaussian distribution
with the exponential by director okay
you have I see so it's a way that you
can combine and chop and change
districts exactly yeah exactly
the idea is distributions are sources of
stochasticity or randomness and by jekt
ours are deterministic
transformations of those sources of
randomness there's a kind of dichotomy
they're interesting so just taking a
step back for a second from and just
think why is probabilistic programming
your view why is it necessary and what
what does it bring to the table for
development yeah that's a exciting
question so I like to think of
probabilistic programming as being
formalism for reasoning under
uncertainty okay I like that definition
because I can't think of something that
doesn't fit into that definition those
are the best definitions those are the
best definitions science is you make a
hypothesis you verify that hypothesis
with some empirical study but we might
imagine that the third time you ran your
experiment you got a totally different
result it doesn't mean your hypothesis
is wrong it just means that guess what
nature has randomness so having a way to
make it easy to specify uncertainty and
to perform inference and reasoning on it
means that you can sort of tease out the
signal from the noise okay and doing so
in a formal way keeps you honest right
it means that you can't claim I have a
cure for cancer when in fact it works
only one out of a million got it got it
and also for safety reasons right if
you're doing something like absolutely
green light absolutely suppose you have
them suppose you have a machine learning
algorithm trying to detect if it's a
green light if you were only two percent
sure that it's a green light you
probably should stop exactly or don't go
don't go yeah but then I gave it's 97
percent or if it's 90 so that's the
question like you like the nice thing
about having some notion of uncertainty
in your predictive system is you can
weigh the opportunity cost of action or
inaction and it might seem like a lot of
this is very theoretical and very
mathematical but it's also very concrete
right as soon as you build your first
neural network as soon as you build a
neuron you know that's entirely based on
probability yeah it's interesting you
say it's theoretical because machine
learning likes to boil things down to
accuracy I have some algorithm and it
achieves 97% accuracy on detecting
handwritten digits that's not how the
real world actually
the whole point of having a machine
learn model is you're going to make some
choice and so I think it's immensely
practical to have a system that tells
you look this is the choice I think you
should make but this is how much you
should trust me that it seems like
actually not theoretical that seems like
what we all really want or maybe even
need and so if you take a probabilistic
first viewpoint you can start to design
systems that have this capability right
and and to be able to get that
probability you need like things like
loss functions and you plug the
distribution into that so by having tons
of distributions exactly the ability to
change see what works exactly and not
just tons of distributions but ways to
change them while preserving the
computational properties and that's the
by actors and that's the by actor
exactly
cool so say I'm getting started in
machine learning and you know even
getting started in machine learning is a
hard enough thing but what advice would
you give me to be able to optimize what
I'm doing what any good pointers that
you yeah yeah so I do not have neural
network as a background but I like to
think of neural nets although this may
be heretical as very fancy forms of
regression mm-hmm regression is a
relatively old actually an old idea in
statistics and relatively old in machine
learning in which you try to predict a
response given some evidence that's what
a neural net does too so I think to get
started we have this thing called
trainable distributions which basically
take a tensor that is your evidence and
it gives you back a distribution from
which you could plug in say minus
distribution dot log prob of your labels
as a loss into an optimizer cool so it
you know the examples we have there show
just how easy it is to do regression in
tensorflow and the nice part about that
is as you add more sophistication
different layers in tensorflow parlance
with different activations nonlinear
functions of the output of those layers
now you can get state-of-the-art machine
learning sophistication cool cool and
obviously if I'm going beyond beginning
and I want to start getting into
distributions what I mean I guess read
your paper would be good advice yes so
the white paper has a lot of examples
that show
just the power of this really simple
idea and those are my favorite kinds of
ideas because I actually get them that's
always the best kind that's the best
kind of idea no point having an idea
that nobody understands especially me
well thank you so much Josh this has
been so much fun and it's like really
really digging down into stuff that I
formerly really didn't understand but
having read your paper and having played
with some of this stuff I can really
begin to see what this is so thanks very
much Josh that was fascinating and thank
you too for watching this episode of
coffee with a Googler if you have any
questions for me or if you have any
questions for Josh please leave them in
the comments below and don't forget to
hit that subscribe button oh and by the
way be sure to check out our brand new
tensorflow channel on youtube thank you