JASON MAYES: Hey, everyone.

My name is Jason Mayes.

I'm a developer advocate within the TensorFlow team

here at Google.

And today, I've got Meghna with me from the TensorFlow Lite

Micro team to talk more about what they

produced in the recent past.

So what have you made, and what is TensorFlow Lite?

MEGHNA NATRAJ: Oh, yeah.

Thanks for the introduction.

So I'm Meghna and I work on the TensorFlow

Lite for microcontroller steam at Google.

So my team basically has built a platform

that helps you to run machine learning models on the edge.

So that-- what that implies is you can try and build

machine learning models using [INAUDIBLE] TensorFlow,

and these are optimized to run on really tiny devices

like Spark Financial that we have

here or on [INAUDIBLE] and so on, and other platforms

that we support.

JASON MAYES: So what are the advantages of TF Micro?

MEGHNA NATRAJ: Yeah, so there are a lot of advantages,

one being that it's super tiny and lightweight so that it

can be a part of a lot of other appliances for doing machine

learning.

So it could be like a refrigerator

or a washing machine.

It's countless.

JASON MAYES: Very flexible.

MEGHNA NATRAJ: Yes.

And the second thing is that it needs no network connectivity.

So it runs without any Wi-Fi or any internet.

And that's because the model is completely running on device.

JASON MAYES: And that's great if in a remote location

like a field or something like this.

MEGHNA NATRAJ: Yeah.

So it has great applications there.

The third one is the fact that it's completely secure.

So there are a lot of security concerns with machine learning

models being used for various other applications,

and in this way, you can assure that the model is completely

running on device and the data is not being transmitted

anywhere else.

JASON MAYES: So can we see a demo in action?

MEGHNA NATRAJ: Yeah.

So as you can see here, the right most LED

is orange in color, which implies that it's detecting

that it's not a person.

And as I move it towards you, the LED

would turn yellow, so the one on the left

if it has just turned yellow, which

implies that it's detected that you're a person.

JASON MAYES: Awesome, great demo.

So how can people get started with this

if they want to at home?

MEGHNA NATRAJ: So we do have a GitHub page and a website

that you can check out.

So it's TensorFlow Lite for Microcontrollers.

Yeah.

That should be a great place to get started.

JASON MAYES: Awesome.

Well, thank you very much for the information today

and happy hacking to those at home.

Hey, everyone.

We're back with Pete Warden from the TensorFlow Lite Micro team,

and we've got a super magical demo for you today.

So tell us more, Pete.

What is this?

PETE WARDEN: So this is a magic wand believe it or not.

JASON MAYES: Excellent.

PETE WARDEN: And if you want to see what it does,

I'm going to try and demonstrate here by doing a W gesture,

and we should see that reflected on the screen.

And there we go.

JASON MAYES: Look at that, a beautiful W on the screen.

PETE WARDEN: And then we try a slope.

JASON MAYES: Uh huh.

And sure enough, we got the slope coming up as well.

PETE WARDEN: Perfect.

JASON MAYES: Awesome.

Cool.

So this is really awesome demo, but how did you

actually create it?

PETE WARDEN: Well, you might be able to tell

by the awesome craftsmanship here

that I actually used some masking tape

to attach this nano 33 BLE sense board from Arduino.

And it actually contains a tiny accelerometer,

which is how it's actually able to recognize the [INAUDIBLE]..

JASON MAYES: I see.

Yes.

Yeah, very cool.

And how much of these go for?

PETE WARDEN: This is about $30.

JASON MAYES: So fairly low cost a venture.

PETE WARDEN: Fairly low cost and you actually

don't need a magic wand.

You can use a stick.

JASON MAYES: Anything is usable.

It's perfect.

Awesome.

So how can we get started if you want to do this at home?

PETE WARDEN: So the nice thing is

we've been up to work with the Arduino community

and get TensorFlow Lite as an official library in the Arduino

ID.

So you can just go to the Manage library's entry and the menu

and get TensorFlow Lite for yourself.

JASON MAYES: So it's pretty much just plug and play then?

PETE WARDEN: Exactly.

JASON MAYES: Thank you very much, Pete.