TensorFlow Dev Summit 2020 基調講演 (TensorFlow Dev Summit 2020 Keynote)

字幕表動画を再生する

[MUSIC PLAYING]
MEGAN KACHOLIA: Hi, everyone.
Welcome to the 2020 TensorFlow Developer Summit Livestream.
I'm Megan Kacholia, VP of Engineering for TensorFlow.
Thanks for tuning into our fourth annual developer
Summit and our first ever virtual event.
With the recent developments of the coronavirus,
we're wishing all of you good health, safety, and well-being.
While we can't meet in person, we're
hoping the Dev Summit is more accessible than ever
to all of you.
We have a lot of great talks along
with exciting announcements, so let's get started.
When we first opensourced TensorFlow,
our goal was to give everyone a platform
to build AI to solve real world problems.
I'd like to share an example of one of those people.
Irwin is a radiologist in the Philippines
and no stranger to bone fracture images like the ones
that you see here.
He's a self-proclaimed AI enthusiast
and wanted to learn how AI could be applied to radiology
but was discouraged because he didn't have a computer science
background.
But then he discovered TensorFlow.js,
which allowed him to build this machine learning application
that could classify bone fracture images.
Now, he hopes to inspire his fellow radiologists
to actively participate in building AI to,
ultimately, help their patients.
And Irwin is not alone.
TensorFlow has been downloaded millions of times
with new stories like Irwin's popping up every day.
And it's a testament to your hard work and contributions
to making TensorFlow what it is today.
So on behalf of the team, I want to say a big thank you
to everyone in our community.
Taking a look back, 2019 was an incredible year for TensorFlow.
We certainly accomplished a lot together.
We kicked off the year with our Dev Summit,
launched several new libraries and online educational courses,
hosted our first Google Summer of Code,
went to 11 different cities for the TensorFlow roadshow,
and hosted the first TensorFlow World last fall.
2019 was also a very special year for TensorFlow
because we launched version 2.0.
It was an important milestone for the platform
because we looked at TensorFlow end to end
and asked ourselves, how can we make it easy to use?
Some of the changes were simplifying the API,
settling on Keras and eager execution,
and enabling production to more devices.
The community really took the changes to heart,
and we've been amazed by what the community has built.
Here are some great examples from winners of our 2.0 Dev
Post Challenge like Disaster Watch, a crisis mapping
platform that aggregates data and predicts
physical constraints caused by a natural disaster,
or DeepPavlov, an NLP library for dialog systems.
And like always, you told us what
you liked about the latest version but,
more importantly, what you wanted to see improved.
Your feedback has been loud and clear.
You told us that building models is easier
but that performance can be improved.
You also are excited about the changes.
But migrating your 1.x system to 2.0 is hard.
We heard you.
And that's why we're excited to share the latest
version, TensorFlow 2.2.
We're building off of the momentum from 2.0 last year.
You've told us speed and performance is important.
That's why we've established a new baseline
so we can measure performance in a more structured way.
For people who have had trouble migrating to 2,
we're making the rest of the ecosystem
compatible so your favorite libraries
and models work with 2.x.
Finally, we're committed to that 2.x core library.
So we won't be making any major changes.
But the latest version is only part
of what we'd like to talk about today.
Today, we want to spend the time talking about the TensorFlow
ecosystem.
You've told us that a big reason why you love TensorFlow
is the ecosystem.
It's made up of libraries and extensions
to help you accomplish your end and end ML goals.
Whether it's to do cutting edge research
or apply ML in the real world, there is a tool for everyone.
If you're a researcher, the ecosystem
gives you control and flexibility
for experimentation.
For applied ML engineers or data scientists,
you get tools that help your models have real world impact.
Finally, there are libraries in the ecosystem that
can help create better AI experiences for your users,
no matter where they are.
All of this is underscored by what all of you, the community,
bring to the ecosystem and our common goal of building AI
responsibly.
We'll touch upon all of these areas today.
Let's start first with talking about the TensorFlow
ecosystem for research.
TensorFlow is being used to push the state of the art of machine
learning in many different sub fields.
For example, natural language processing
is an area where we've seen TensorFlow really
help push the limits in model architecture.
The T5 model on the left uses the latest
in transfer learning to convert every language problem
into a text-to-text format.
The model has over 11 billion parameters
and was trained off of the colossal clean crawled corpus
data set.
Meanwhile, Meena, the conversational model
on the right, has over 2.6 billion parameters
and is flexible enough to respond sensibly
to conversational context.
Both of these models were built using TensorFlow.
And these are just a couple examples
of what TensorFlow is being used for in research.
There are hundreds of papers and posters
that were presented at NeurIPS last year that used TensorFlow.
We're really impressed with the research
produced with TensorFlow every day
at Google and outside of it.
And we're humbled that you trust TensorFlow
with your experiments, so thank you.
But we're always looking for ways
to make your experience better.
I want to highlight a few features
in the ecosystem that will help you in your experiments.
First, we've gotten a lot of positive feedback
from researchers on TensorBoard.dev,
a tool we launched last year that lets you upload and share
your experiment results by URL.
The URL allows for quickly visualizing hyper parameter
sweeps.
At NeurIPS, we were happy to see papers starting
to cite TensorBoard.dev URLs so that other researchers could
share experiment results.
Second, we're excited to introduce a new performance
profiler toolset in TensorBoard that
provides consistent monitoring of model performance.
We're hoping researchers will love the toolset because it
gives you a clear view of how your model is performing,
including in-depth debugging guidance.
You'll get to hear more about TensorBoard.dev
and the new profiler from Gal and [? Schumann's ?] talks
later today.
Researchers have also told us that the changes in 2.x
make it easy for them to implement new ideas, changes
like eager execution in the core.
It supports numpy arrays directly,
just like all the packages in the py data
ecosystem you know and love.
The tf.data pipelines we rolled out are all reusable.
Make sure you don't miss Rohan's tf.data talk today
for the latest updates.
And TensorFlow data sets are ready right out of the box.
Many of the data sets you'll find
were added by our Google Summer of Code students.
So I want to thank all of them for contributing.
This is a great example of how the TF ecosystem is
powered by the community.
Finally, I want to round out the TensorFlow ecosystem
for research by highlighting some
of the add ons and extensions that researchers love.
Libraries like TF probability and TF agents
work with the latest version.
And experimental libraries like JAX from Google Research
are composable all with TensorFlow
like using TensorFlow data pipelines to input data
into JAX.
But TensorFlow has never just been
about pushing the state of the art in deep learning.
A model is only as good as the impact
it has in the real world.
This is one of TensorFlow's core strengths.
It has helped AI scale to billions of users.
We've seen incredible ML applications
being built with TensorFlow.
We're really humbled by all the companies, big and small,
who trust TensorFlow with their ML workloads.
Going from an idea to having your AI create real world
impact can be hard.
But our users rely on TensorFlow to help them accomplish this.
That's because the TensorFlow ecosystem
is built to fit your needs.
It makes having to go from training
to deployment less of a hassle because you have the libraries
and resources all in one platform.
There's no switching costs involved.
I want to highlight a few new things that will help
you get to production faster.
First, you've told us that you love
working with Keras in TensorFlow because it's easy to build
and train custom models.
So we're committed to keeping tf.keras,
the default high level API.
But if you're not looking to build models from scratch,
TensorFlow Hub hosts all the ready to use pre-trained models
in the ecosystem.
There are more than 1,000 models available in TF Hub
with documentation, code snippets, demos,
and interactive collabs, all ready to be used.
When you're ready to put your model into production,
you can build production ready ML pipelines
in TensorFlow Extended to make sure your ML engineering just
works, from data validation to ML metadata tracking.
And today, I'm very excited to announce
that using TensorFlow in production
is getting even easier with an exciting launch, Google Cloud
AL Platform pipelines.
We've partnered with Google Cloud
to make it easy to build end-to-end production
pipelines using Kubeflow and TensorFlow Extended, hosted
by Google Cloud.
Cloud AI platform pipelines are available today
in your Google Cloud console.
And if you're running TensorFlow on Google Cloud,
TensorFlow Enterprise, which we announced last year
at TF World, gives you the long term support and the enterprise
scale that you need.
Finally, you can train and deploy
your models and pipelines on custom hardware specifically
designed for AL workloads, cloud TPUs.
In the latest version, TensorFlow
is now optimized for cloud TPUs using Keras.
This means the same API you started
with now helps you scale to petraflops of TPU compute.
All of these libraries are within the TensorFlow
ecosystem, are 2.2 compatible, and help
you scale so your ML application can reach your users.
But for AL to have that kind of impact,
it needs to be where your users are,
which means getting your models on device.
Now, we all know this requires working
in some constraints like low latency,
working with poor network connectivity,
all while trying to preserve privacy.
You can do all of this by using tools within the TensorFlow
ecosystem, like TensorFlow Lite, which can help make your models
run as fast as possible, whether it's on CPUs, GPUs,
DSPs, or other accelerators.
Here's an example of how we've optimized
performance for MobileNet V1 from May last year to today.
It's a big reduction in latency and something
you get right out of the box with TF Lite.
We're also adding Android Studio integration
so you can deployment models easily.
Just simply drag and drop in Android Studio
and automatically generate the Java classes for the TF Lite
model with just a few clicks.
When network connectivity is a problem
and you need these power intensive models to work
while still offline, you can convert
them to run better on device using TensorFlow Lite.
In the latest version, we rebuilt the TF Lite converter
from the ground up to provide support
for more models, more intuitive error messages
when conversions fail, and support for control flow
operations.
The browser has become an exciting place
for interactive ML.
And TensorFlow.js is allowing JavaScript and web developers
to build some incredible applications.
There's some exciting new models that are now supported
like FaceMesh and MobileBERT.
HuggingFace introduced a new NPM package
for tf.js, which allows you to do question answering directly
in Node.js.
Finally, the new WebAssembly backend
is available for improved CPU performance.
The next few years will see an explosion
of platforms and devices for machine learning.
And the industry needs a way to keep up.
MLIR is a solution to this rapidly changing landscape.
It's compiler infrastructure for TF and other frameworks.
And it's backed by 95% of the world's hardware accelerator
manufacturers and is helping to move the industry forward.
We see how important infrastructure like MLIR
is to the future of ML, which is why
we're investing in the future of TensorFlow's
own infrastructure.
The new TensorFlow runtime is something
you won't be exposed to as a developer or researcher.
But it will be working under the covers
to give you the best performance possible across a wide variety
of domains specific hardware.
We're planning on integrating the new runtime
later this year.
But you'll hear more from Mingsheng later today.
So to recap everything you've seen so far,
whether you're pushing the state of the art in research,
applying ML to real world problems,
or looking to deploy AI wherever your users are,
there is a tool for you in the TensorFlow ecosystem.
Now, I'd like to invite Manasi on stage
to talk about how the ecosystem is
helping empower responsible AI.
Thank you.
MANASI JOSHI: Thank you.
Thanks, Megan.
Hi, everyone.
My name is Manasi Joshi.
And I'm an engineering director on TensorFlow team.
As Megan mentioned and you saw, TensorFlow ecosystem
is composed of a number of useful libraries and tools
that are useful for a diverse set of use cases,
whether they're coming from ML researchers or practitioners
alike.
However, the field of ML and AI is raising the question
whether we are building systems in the most
inclusive and secure way.
I'm here to tell you how TensorFlow ecosystem empowers
its users to build systems responsibly
and, moreover, what type of tools and resources
are available to our users to accomplish those goals.
Before we deep dive into the details of what TensorFlow
has to offer its users, let's take a step back and define
what we mean by responsible AL.
As we know, that machine learning
has tremendous power for solving lots of challenging real world
problems.
However, we have to do this responsibly.
Now, to us, in TensorFlow, the way we define responsible
AI is based on a five pillar strategy.
Number one, general recommended practices for AI--
this is all about reliability, all the way from making sure
that your model is not over fitting to your training data--
it is more generalized than that--
making sure you are aware of limitations of your training
data when it comes to different feature
distribution [INAUDIBLE],, for example, ensuring
that the model outputs are robust
when the training data gets perturbed,
ensuring you're not using only a single metric across all
your models to determine its quality
because different metrics matter to different context,
how your model is used for promotion, demotion, filtering,
ranking, so on and so forth.
The second principle, fairness--
fairness is a fairly evolving thing in AI.
For us, we define it as not to create or reinforce
unwanted bias.
Fairness can be extremely subjective, can
be context sensitive.
And it is a socio-technical challenge.
Third, interpretability-- interpretability
is all about understanding the mechanics
behind a model's prediction, ensuring that you understand
what features really matter to the final output, which
features were important, which features were not.
Fourth, privacy-- for us, this is all about
taking into account sensitivity of your training data
and features.
And fifth is security.
In the context of ML, security really
means that you understand [INAUDIBLE] liabilities
in your system and the threat models that are associated.
Now, for a typical user of TensorFlow,
this is how the overall developer workflow looks like.
You start by defining a specific goal and an objective
for why you want to build a system.
Then you go about gathering relevant data
for training your model.
As we understand, data is gold for machine
learning module training.
And so you have to prepare the data well.
You have to transform it.
You have to cleanse it sometimes.
And then, once your data is ready,
you go about training your model.
Once the model is built-- it's converged--
you then go about deploying the model in production systems
that want to make use of ML.
Deployment phase is not a one time task.
You have to continuously keep iterating in an ML workflow
and improving the quality of the model.
Now, along this developer workflow, there are many,
many different moments at which you, as a modeler,
needs to be asking all of these questions,
questions like, who is the audience for my machine
learning model?
Who are the stakeholders?
And what are the individual objectives
for the stakeholders?
Going onto the data side of it, is my data really representing
real world biases or distribution skews?
And do I understand those limitations?
Am I allowed to use certain features
in a privacy preserving way?
Or are they are just simply not available due to constraints?
Then onto the training side of it,
do I understand implications of the data
on model outputs or not?
Am I doing deployments very blindly or?
Am I being a little bit mindful about deploying
only reliable and inclusive models?
And finally, when we talk about iterative workflow,
do I understand complex feedback loops that
could be present in my modeling workflow?
Now, along all of these questions,
I'm happy to tell you that TensorFlow ecosystem has
a few set of tools which could be
helpful to answer some of them.
I'm not going to go through everything
here but to just give you a few examples,
starting with Fairness Indicators.
It's a very effective way by which
you can evaluate your model's performance
across many different subgroups in a confidence
interval powered way such that you
can evaluate simple but effective fairness
metrics for your models.
We have What-If tool that gives you
the notion of interpreting the model's output based
on the features and changing those features to see
the changes in the model's output.
It has very compelling textual as well as visual information
associated with your data.
And then finally, TensorFlow Federated
is a TensorFlow 2.x compatible library
that helps you train your models with data
that's available on device.
Cat and Miguel have a talk later today
that dives deep into fairness and privacy pillars
of the responsible AL strategy.
Be sure not to miss it.
We are excited to work on this important part of TensorFlow
ecosystem with all of you, the TensorFlow community.
And now, to talk more about the community,
I would like to turn it over to Kemal.
Thank you.
KEMAL MOUJAHID: Thank you, Manasi.
Hi, everyone.
My name is Kemal, and I'm the product director
for TensorFlow.
So you've heard a lot of from Megan and Manasi
about our latest innovations.
Now, I'm going to talk about the most important part of what
we're building.
And that's the community.
And I want to begin by thanking all of you.
Your feedback, your contributions, what you build--
this is what makes all of this possible.
We have an incredible global community.
We love hearing from you.
And we really appreciate everyone
that came out to a roadshow of-- or TensorFlow World last year.
And going into 2020, I want to take some time
to highlight more opportunities to get involved
in the community and new resources
to help you all succeed.
Let's start with ways you can get involved locally.
One great way to connect is to join a TensorFlow user group.
These grassroots communities start organically.
And we now have 73 of them globally.
We launched our first two in Latin America
after the roadshow in San Paolo.
And now, we've expanded our presence in Europe.
The Korea group is the biggest with 46,000 members.
And China has user groups in 16 cities.
I'm sure this map can have a lot more dots.
So if you want to start a user group, please reach out.
And we'll help you get started.
Another way to get involved are the special interest groups,
or SIGs.
To help you build areas of TensorFlow
that you care the most about, we will have 12 SIGs
with SIG graphics being our latest addition starting
at the end of the month.
Most SIGs are led by members of the open source
community such as our fantastic Google Developer Experts.
We love our GDEs.
We now have 148 of them.
And I want to take a moment to recognize all of them.
They give tech talks, organize workshops and doc sprints.
And I want to give a special shout out
to [? Hasan, ?] pictured above, who organizes
a doc sprint in Seoul.
They reviewed several PRs and wrote hundreds
of comments in five hours.
Again, GDEs are amazing.
So please let us know if you're interested in becoming one.
OK, so TensorFlow user groups, SIGs, and GDEs
are great ways to get involved.
But we all love a little competition.
And we all love Kaggle.
As Kaggle now supports 2.x, we've
had over 1,000 teams enrolled in our last competition.
I want to give a special shout out to our 2.0 prize
winner, Deep Thought.
And speaking of competition, we saw great projects
in our last DevPost challenge, including Psychopathology
Assistant, an intelligent assistant that
tracks patient's responses during face-to-face and remote
sessions, and Everyone Dance Faster, an everybody dance now
video generation library using [? HTTPU ?] and TensorFlow 2.0.
Thank you to everyone who participated.
And today, we have a new challenge.
Manasi spoke earlier about how TensorFlow
can help empower all users to build AI systems responsibly.
So we want to challenge you to create something great
with TensorFlow 2.2 and something that
has the AI principles at heart.
We can't wait to see what you build.
So another area that we're investing in a lot
is education, starting with supporting
our younger community members.
For the first time, we participated in Google coding.
And it was a success.
We were very impressed by the students.
And we want to thank all the awesome mentors who
made this possible.
We hope someday to see the students
at our Summer of Code program.
I love Summer of Code.
It's an awesome opportunity for students
to work with TensorFlow engineers.
We saw amazing projects.
In fact, one of the students worked on data visualization
for Swift, which is still being used today by our team.
So I'm happy announce we're doing it again this summer.
And we're excited to see what new projects students will
work on.
Programs like this are key to the growth of the developer
community.
So please visit the Summer of Code website
to learn more and apply.
We also want to help produce great educational content,
starting with our machine learning crash course,
a great resource for beginners.
So today, we launched an updated version of the course.
Our [INAUDIBLE] team completely revamped
the programming exercises using Keras 2.0
and made them much simpler in the process.
Go check it out on this link.
And we want to provide resources at every stage of learning.
At a university level, we want to empower educators
and support them as they design, develop, and teach
machine learning courses.
Last year, we supported Georgia Tech, University of Hong Kong,
Pace University, and many others.
And this year, we have a commitment
to fund schools from underrepresented communities
in AI, historically black and Latinx colleges
and universities.
So if you're a faculty and you want to teach ML,
please reach out.
And we also want to help people self-study.
That's why we partner with deeplearning.ai
to give people access to great educational material.
To date, over 200,000 people have enrolled in our courses.
The Data and Deployment course is a great specialization
course that covers TensorFlow GS, TensorFlow Lite, TensorFlow
Dataset, and more advanced scenarios.
This is a great option for people
who are really looking to build their ML coding skills.
And you could audit it for free.
And there's more.
Imperial College London just released a
getting started with TensorFlow course on Coursera.
This course was created in part by the TensorFlow funding
I mentioned earlier.
And we're super happy to see this.
So you're taking all these courses.
You're becoming better at ML.
But how do you show your expertise to the world?
This is why I'm excited to announce
the launch of the TensorFlow certificate program,
an assessment created by the TensorFlow team,
covering topics such as text classification using
NLP to build spam filters, computer vision using
CNN to do image recognition, sequences and prediction.
By passing this foundational certification,
you'll be able to share your expertise with the world
and display your certificate badge on LinkedIn, GitHub,
or the TensorFlow certificate network.
And to widen access to people of diverse backgrounds
and experiences, we're excited to offer
a limited number of stipends for covering the certification
costs.
You can find out more at TensorFlow.org/certificate.
So a lot of things to do, and I want
to thank you, again, for making the TensorFlow community so
awesome.
As you've seen, the TensorFlow ecosystem
is having incredible impact in the world today.
And what it really comes down to is
how AI is helping make people's lives better.
That's really what inspires us, as a team,
to build all these amazing tools.
So I'd like to end by sharing one final story.
[VIDEO PLAYBACK]
[END PLAYBACK]
KEMAL MOUJAHID: That's just amazing
and incredibly inspiring.
When I see something like this, it
makes me very proud to be building TensorFlow.
So go build amazing things, and we'll be there to help.
And with that, I'll pass it on to Paige to kick off our day.
Thank you.
[MUSIC PLAYING]