TensorFlow: Machine Learning for Everyone

ZAK STONE: It’s been
just over a year since we open-sourced
TensorFlow, and we’ve been thrilled to see
the adoption by the community, and the pace of development
both here at Google and all around the world. JEFF DEAN: TensorFlow is
really the primary tool that we’re using for
a lot of our machine learning work in
all of our products. Towards the end of
last year, we actually rolled out a completely
new translation system that was based on
deep neural nets. In Gmail, we were actually able
to roll out a TensorFlow model that, by understanding the
context of the message you just received, we can
predict likely replies, and this is a feature
we call Smart Reply. LILY PENG: Diabetic retinopathy
is the fastest growing cause of blindness. It’s a complication of diabetes. We gathered a very
large data set and had doctors
grade the images, and then we, using TensorFlow,
trained a neural net that does a pretty good job
of predicting whether or not there is a diabetic
retinopathy in the image. DOUG ECK: Can we use something
like TensorFlow to make music, to make art, and to allow
us to communicate better with each other? With TensorFlow, we’re
able to think abstractly, almost at a level of, like,
improvisation with machine learning. We’re able to try new
things, to chunk models together in ways that
were impossible before we had that kind of expressivity. AMANDA HODGSON: Dugongs are classed
as vulnerable to extinction globally. So we do a lot of aerial
surveys using drones. Then once you’ve done a
survey of a really large area, you end up with tens, if
not hundreds of thousands of photos. The goal was to find a way to
automate that whole process. And that’s where we’ve
been using TensorFlow. ZAK STONE: One of
the things that we’ve been focusing on this year
with TensorFlow is performance. We’ve been especially
excited to release support for distributed training. MEGAN KACHOLIA: We want to make
it easier for people to use, so they don’t have
to necessarily know all of the underlying
internals in order to get the distributed
performance the best it can be. [INAUDIBLE] is something that
can compile down TensorFlow. Maybe you want to compile
your graph ahead of time and then get it down to
something much more compact, in terms of memory size. So that that way, you can
easily load it and execute it on something that might not
have as much storage space, like a mobile phone or some
other portable smaller device. RICK MAULE: When we introduced
the hexagon vector extensions, what we had in mind was
enhancing user experiences with imaging features. ERICH PLONDKE: So
the TensorFlow team said that you only needed
low precision multiplies to be able to execute these
neural networks efficiently. So we did some tests and on
the same graph, Inception_v3, we were eight times faster
and four times lower power than running on the CPUs. RICK MAULE: TensorFlow
is great to work with, easy to work with,
lots of capability. So our engineering teams
and their engineering teams working together, we were able
to do something very exciting. This is just the
beginning of what will end up being a long
evolution of some great things we can do with machine
learning and image processing. JOSH GORDON: In addition
to sharing TensorFlow, Google has also
shared a ecosystem of tools, which
contains everything you need to go all the way
from research to production. One such tool is
TensorFlow serving, and this is a open-source,
high-performance serving solution. Another great tool, which
is actually quite beautiful, is the embedding visualizer. And you can use the
embedding visualizer to interactively explore
high-dimensional data sets. On the education
side, General Assembly has done great work
teaching TensorFlow. NEHEMIAH LOURY: For
my final project, I was really interested in
doing lyrics generation. And TensorFlow was a
really great match for that because it allowed me to build
out and utilize the models that I needed to be successful. ZAK STONE: The
TensorFlow community is thriving around
the world, and we’re excited about as many people
as possible being part of it. TensorFlow is an open-source
project for everyone. We’re looking forward
to building this into something even better and
more useful and more powerful, in collaboration with the
whole worldwide community.

Leave a Reply

Your email address will not be published. Required fields are marked *