PETE WARDEN: So thanks so much, Raziel.

And I'm real excited to be here to talk about a new project

that I think is pretty cool.

So TensorFlow Lite for microcontrollers--

what's that all about?

So this all comes back to when I actually first joined Google

back in 2014.

And as you can imagine, there were a whole bunch

of internal projects that I didn't actually

know about, as a member of the public, that

sort of blew my mind.

But one in particular came about when I actually

spoke to Raziel for the first time, and he explained.

And he was on the speech team at the time working

with Alex, who you just saw.

And he explained that they use neural network models

of only 13 kilobytes in size.

At that time I only really had experience with image networks.

And the very smallest of them was still

like multiple megabytes.

So this idea of having a 13-kilobyte model

was just amazing for me.

And what amazed me even more was when

he told me why these models had to be so small.

They needed to run them on these DSPs and other embedded chips

in smartphones.

So Android could listen out for wake words, like "Hey, Google,"

while the main CPU was powered off to save the battery.

These microcontrollers often only

had tens of kilobytes of RAM and flash storage.

So they simply couldn't fit anything larger.

They also couldn't rely on cloud connectivity

because the amount of power that would have been drained just

keeping a radio connection alive to send data over

would have just been prohibitive.

So that really struck me, that conversation

and the continued work that we did with the speech team,

because they had so much experience

doing all sorts of different approaches with speech.

They'd spent a lot of time and a lot of energy experimenting.

And even within the tough constraints

of these embedded devices, neural networks

were better than any of the traditional methods they used.

So I was left wondering if they'd

be really useful for other embedded sensor applications

as well.

And it left me really wanting to see if we could actually

build support for these kind of devices into TensorFlow itself,

so that more people could actually get access.

At the time, only people in the speech community

really knew about the groundbreaking work that

was being done, so I really wanted

to share it a lot more widely.

So [LAUGHS] today, I'm pleased to announce

that we are releasing the first experimental support

for embedded platforms in TensorFlow Lite.

And to show you what I mean, here

is a demonstration board that I actually have in my pocket.

And this is a prototype of a development

board built by SparkFun.

And it has a Cortex-M4 processor with 384 kilobytes of RAM

and a whole megabyte of flash storage.

And it was built by Ambiq to be extremely low power,

drawing less than one milliwatt in a lot of cases.

So it's able to run on a single coin battery like this

for many days, potentially.

And I'm actually going to take my life in my hands

now by trying a live demo.

[LAUGHS] So let us see if this is actually--

it's going to be extremely hard to see,

unless we dim the lights.

There we go.

So what I'm going to be doing here

is, by saying a particular word, and see if it actually lights

up the little yellow light.

You can see the blue LED flashing.

That's just telling me that it's running [INAUDIBLE]..

So if I try saying, yes.

Yes.

[LAUGHS] Yes.

[LAUGHS] I knew I was taking my life into my hands here.

[LAUGHTER]

Yes.

There we go.

[LAUGHS]

[APPLAUSE]

So I'm going to quickly move that out of the spotlight.

[LAUGHS] So as you can see, it's still far from perfect.

[LAUGHS] But it is managing to do

a job of recognizing when I say the word,

and not lighting up when there's unrelated conversations.

So why is this useful?

Well first, this is running entirely locally

on the embedded chip.

So we don't need to have any internet connection.

So it's a good, useful first component of a voice interface

system.

And the model itself isn't quite 13 kilobytes,

but it is down to 20 kilobytes.

So it only takes up 20 kilobytes of flash storage

on this device.

And the footprint of the TensorFlow Lite

code for microcontrollers is only another 25 kilobytes.

And it only needs about 30 kilobytes of RAM

available to operate.

So it's within the capabilities of a lot

of different embedded devices.

Secondly, this is all open source.

So you can actually grab the code yourself

and build it yourself.

And you can modify it.

I'm showing you here on this particular platform,

but it actually works on a whole bunch

of different embedded chips.

And we really want to see lots more supported,

so we're keen to work with the community

on collaborating to get more devices supported.

You can also train your own model.

Just something that recognizes yes isn't all that useful.

But the key thing is that this comes with a [INAUDIBLE]

that you can use to actually train your own models.

And it also comes with a data set

of 100,000 utterances of about 20 common words

that you use as your training set.

And that first link there, the aiyprojects one,

if you could actually go to that link

and contribute your voice to the open data set,

it should actually increase the size

and the quality of the data set that we can actually

make available.

So that would be awesome.

And you can actually use the same approach

to do a lot of different audio recognition

to recognize different kinds of sounds,

and even start to use it for similar signal processing

problems, like things like predictive maintenance.

So how can you try this out for yourself?

If you're in the audience here, at the end of today,

you will find that you get a gift box.

And you actually have one of these in there.

[APPLAUSE]

[LAUGHS]

And all you should need to do is remove the little tab

between the battery, and it should automatically boot up,

pre-flashed, with this yes example.

[LAUGHTER]

So you can try it out for yourself,

and let me know how it goes.

Just say yes to TensorFlow Lite is the--

[LAUGHTER]

And we also include all the cables,

SO you should be able to just program it

yourself through the serial port.

Now these are the first 700 boards ever built,

so there is a wiring issue.

So it will drain the battery.

It won't last.

It would last more like hours than days.

But that will actually, knock on wood,

be fixed in the final product that's shipping.

And you should be able to develop

with these in the exact same way that you

will with the final shipping product.

And if you're watching at home, you

can pre-order one of these form SmartFun

right now for, I think, it's $15.

And you'll also find lots of other instructions

for other platforms in the documentation.

So we are trying to support as many

of the modern microcontrollers that are out there that people

are using as possible.

And we welcome collaboration with everybody

across the community to help unlock all of the creativity

that I know is out there.

And I'm really hoping that I'm going

to be spending a lot of my time over the next few months

reviewing pull requests.

And finally, this is my first hardware project,

so I needed a lot of help from a lot of people

to actually help bring this prototype together,

including the TF Lite team, especially Raziel, Rocky, Dan,

Tim, and Andy.

Alister, Nathan, Owen, and Jim at SparkFun were lifesavers.

We literally got these in our hands middle of the day

yesterday.

[LAUGHTER]

So the fact that they managed to pull it together

is a massive tribute.

And also Scott, Steve, Arpit, and Andre

at Ambiq, who actually designed this process and helped us

get the software going.

And actually a lot of people at Arm as well, including a big

shout out to Neil and Zach.

So this is still a very early experiment,

but I really can't wait to see what people build with this.

And one final note.

I will be around to talk about MCUs

with anybody who's interested at the breakout session on day

two.

So I'm really looking forward to chatting to everyone.

Thank you.

[APPLAUSE]

RAZIEL ALVAREZ: Thanks, Pete.

We really hope that you try this.

I mean, it's the early stages, but you see a huge effort

just to make this happen.

We think that it will be really impactful for everybody.

Now before we go again--

and I promise this is the last thing you hear from me--

I want to welcome June, who's going

to talk about how, by using TensorFlow Lite with the Edge

TPU Delegate are able to train these teachable machines.

[MUSIC PLAYING]

[APPLAUSE]

JUNE TATE-GANS: Thanks, Raziel.

Hi.

My name is June Tate-Gans.

I'm actually one of the lead software engineers inside

of Google's new Coral Group.

And I've been asked to give a talk about the Edge

TPU-based teachable machine demo.

So first, I should tell you what Coral is.

Coral is a platform for products with on-device machine learning

using TensorFlow and TF Lite.

Our first two products are a single-board computer

and a USB stick.

So what is the Edge TPU?

It's a Google-designed ASIC that accelerates inference directly

on the device that it's embedded in.

It's very fast.

It localizes data to the edge, rather than the cloud.

It doesn't require a network connection to run.

And this allows for a whole new range of applications

of machine learning.

So the first product we built is the Coral Dev Board.

This is a single-board computer with a removable SOM.

It runs Linux and Android.

And the SOM itself has a gigabyte

of RAM, a quad-core A53 SoC, Wi-Fi and Bluetooth,

and of course the Edge TPU.

And the second is our Coral accelerator board.

Now, this board is just the Edge TPU connected

via USB-C to whatever development system

you need, be it a Raspberry Pi, or a Linux workstation.

Now, this teachable machine shows off

a form of edge training.

Now traditionally, there's three ways to do edge training.

There's k-nearest neighbors, weight imprinting,

and last layer retraining.

But for this demo, we're actually

using the k-nearest neighbors approach.

So in this animated GIF, you can see that the TPU enables

very high classification rates.