WEBVTT

00:00.000 --> 00:07.000
Thank you so much.

00:07.000 --> 00:09.000
Yeah.

00:09.000 --> 00:10.000
Last talk of the day.

00:10.000 --> 00:14.000
We'll talk about machine learning for very, very small systems,

00:14.000 --> 00:17.000
especially compared to what's all the rage today,

00:17.000 --> 00:19.000
with giant language models, generative AI,

00:19.000 --> 00:20.000
and so on.

00:20.000 --> 00:23.000
This will be completely on different side of the spectrum,

00:23.000 --> 00:27.000
and talking about how we can do machine learning that is useful,

00:27.000 --> 00:30.000
for different kind of tasks at the machine,

00:30.000 --> 00:34.000
to be somewhat size on small microsystems.

00:34.000 --> 00:36.000
So when it is you and them from sound sensing,

00:36.000 --> 00:39.000
we monitor ventilation systems,

00:39.000 --> 00:41.000
such as the ones shown here,

00:41.000 --> 00:43.000
and our customers are operating buildings such as this,

00:43.000 --> 00:46.000
and for us we use machine learning to automatically analyze

00:46.000 --> 00:48.000
to detect anomalies.

00:48.000 --> 00:53.000
And today we'll talk about why machine learning on micro-controllers

00:54.000 --> 00:57.000
is relevant, like how can we do useful things,

00:57.000 --> 00:59.000
which on such small systems?

00:59.000 --> 01:01.000
This niche is called tiny ML,

01:01.000 --> 01:05.000
or edge AI is also getting popular as a term.

01:05.000 --> 01:08.000
We'll talk about the EM Learn project,

01:08.000 --> 01:10.000
the open source project that I maintain,

01:10.000 --> 01:13.000
some projects that has been made,

01:13.000 --> 01:15.000
real world projects with EM Learn,

01:15.000 --> 01:18.000
and we'll cover a quick how through too.

01:18.000 --> 01:25.000
So machine learning on micro-controllers and embedded devices,

01:25.000 --> 01:28.000
what is it used for?

01:28.000 --> 01:32.000
It's primarily used for sensor data analysis.

01:32.000 --> 01:35.000
So we have some sort of sensor node.

01:35.000 --> 01:37.000
We have the sensor itself,

01:37.000 --> 01:40.000
can be camera, experimenter, microphone, radar,

01:40.000 --> 01:43.000
and many lightar, et cetera.

01:43.000 --> 01:46.000
And we want to process that information,

01:46.000 --> 01:49.000
or all that data, and extract the useful information

01:49.000 --> 01:53.000
from that large amount of data.

01:53.000 --> 01:56.000
And only, like for example, directly act on this,

01:56.000 --> 01:58.000
in robotics, for example,

01:58.000 --> 02:01.000
or we might transfer it over to a larger system,

02:01.000 --> 02:02.000
or potentially to a cloud.

02:02.000 --> 02:04.000
But just the relevant information,

02:04.000 --> 02:07.000
not all the useless data.

02:07.000 --> 02:09.000
So this allows some key benefits.

02:09.000 --> 02:12.000
We can make standalone systems that operate

02:12.000 --> 02:15.000
with low latency, guaranteed latency.

02:15.000 --> 02:17.000
Highly power efficient systems that can,

02:17.000 --> 02:21.000
where the sensors can run for multiple years on the single battery.

02:21.000 --> 02:24.000
And we can also make privacy compatible systems,

02:24.000 --> 02:28.000
because we don't transfer all the potentially sensitive data.

02:28.000 --> 02:30.000
And we can also make things low costs,

02:30.000 --> 02:33.000
which is relevant because it enables massive scale

02:33.000 --> 02:36.000
in terms of the number of units we will deploy.

02:36.000 --> 02:39.000
And this is already used in many areas,

02:39.000 --> 02:41.000
for example, in consumer technology,

02:41.000 --> 02:45.000
keyboard spotting, so detecting hay Google, or hay Siri.

02:45.000 --> 02:47.000
And so on, that's using microtroller,

02:47.000 --> 02:49.000
then wakes up the system.

02:49.000 --> 02:51.000
Or if you have a sleep quality tracker,

02:51.000 --> 02:55.000
or a fitness tracker, that will also use this kind of technology.

02:55.000 --> 02:57.000
And in industrial case, for example,

02:57.000 --> 03:01.000
you can use this to track the health of animals.

03:01.000 --> 03:04.000
We use it to track the health of machines.

03:04.000 --> 03:08.000
Or you can just use it for fun projects at home.

03:08.000 --> 03:10.000
For example, you can make a smart doorbell,

03:10.000 --> 03:12.000
which opens the door for your cats,

03:12.000 --> 03:14.000
but only if it's your specific cat,

03:14.000 --> 03:17.000
and not any cat, or you can make a magic wand,

03:17.000 --> 03:19.000
which you can use to make a gesture

03:19.000 --> 03:22.000
to control some of the devices in your home.

03:22.000 --> 03:24.000
And these microtrollers,

03:24.000 --> 03:26.000
they are essentially a full computer,

03:26.000 --> 03:28.000
but very, very small one.

03:28.000 --> 03:32.000
You have, for example, maybe one kilobytes of RAM,

03:32.000 --> 03:35.000
or maybe the largest system has 1000 kilobytes of RAM,

03:35.000 --> 03:38.000
and similar in terms of the program space.

03:38.000 --> 03:40.000
All the machine learning, all the code,

03:40.000 --> 03:43.000
needs to fit in there.

03:43.000 --> 03:46.000
And we're mostly doing inference on the edge.

03:46.000 --> 03:48.000
Not so much learning on the edge.

03:48.000 --> 03:52.000
The price point might be as low as 10 cents,

03:52.000 --> 03:56.000
or as much as 10 dollars for a microtroller.

03:56.000 --> 03:58.000
And this is a huge area.

03:58.000 --> 04:00.000
These like any electronics will have one or more.

04:00.000 --> 04:04.000
Like your car will have 100 microcontrollers these days.

04:04.000 --> 04:08.000
So there are over 20 billion devices shipped each year.

04:08.000 --> 04:12.000
So this is tiny chips, but massive scale.

04:12.000 --> 04:14.000
And this is increasingly available for hobbyists.

04:14.000 --> 04:17.000
So you can program these kind of things with the R.

04:17.000 --> 04:21.000
We know ID, for example, or with micro Python.

04:21.000 --> 04:24.000
But efficiency is key.

04:24.000 --> 04:28.000
So the EMLARM project I started in 2018,

04:28.000 --> 04:31.000
as part of my master in data science,

04:31.000 --> 04:33.000
before I have a background lesson.

04:33.000 --> 04:37.000
In general, I wanted to combine those skills.

04:37.000 --> 04:40.000
And there's two aspects to these kind of systems.

04:40.000 --> 04:42.000
You have the training process.

04:42.000 --> 04:44.000
That's quite standard.

04:44.000 --> 04:45.000
It happens on your computer.

04:45.000 --> 04:48.000
Use familiar tools, such as psychic learn or caras.

04:48.000 --> 04:51.000
And you have a pipeline that fits out of the deploy model.

04:51.000 --> 04:55.000
And then we have the EMLARM project.

04:55.000 --> 04:58.000
It's a Python library that you install.

04:58.000 --> 05:02.000
And you have a one line export to a efficient device.

05:02.000 --> 05:11.000
And then on the device side, you can then deploy this either with EMLARM C library,

05:11.000 --> 05:14.000
which is written in portable C99.

05:14.000 --> 05:20.000
It will run any system that can run C, which is basically any embedded system.

05:20.000 --> 05:23.000
This is the link of Franca of that area.

05:23.000 --> 05:28.000
And it's even run used inside the Linux kernel modules, for example.

05:28.000 --> 05:31.000
Or you can use it with EMLARM micro Python,

05:31.000 --> 05:36.000
which is quite interesting because most people

05:36.000 --> 05:39.000
let the machine learning they know Python primarily.

05:39.000 --> 05:43.000
So this enables to do all the application logic for even this kind of

05:43.000 --> 05:46.000
tiny ML type systems in Python.

05:46.000 --> 05:53.000
And then the EMLARM library is implemented in C and exposing Python APIs.

05:53.000 --> 05:57.000
EMLARM supports the most common tasks,

05:57.000 --> 06:01.000
such as classification, regression, and analytics that are relevant for

06:01.000 --> 06:06.000
these kind of sensor data and better use cases.

06:06.000 --> 06:11.000
Support a range of simple and effective embedded friendly model.

06:11.000 --> 06:12.000
So this is a subset.

06:12.000 --> 06:14.000
So there's no large language model.

06:14.000 --> 06:15.000
There's nothing like that.

06:15.000 --> 06:18.000
But you have, for example, decision trees, random forest,

06:18.000 --> 06:20.000
classification, k-nearest neighbors.

06:20.000 --> 06:23.000
You can do simple and device learning or adaptation.

06:23.000 --> 06:27.000
And also some small neural networks, multi-layer perceptrons,

06:27.000 --> 06:31.000
convolutional neural networks.

06:31.000 --> 06:34.000
And here are some projects you think in learn, of course,

06:34.000 --> 06:35.000
being open source.

06:35.000 --> 06:38.000
Most of the projects out there we will never hear about,

06:38.000 --> 06:41.000
but it's at least referenced in over 40 publications

06:41.000 --> 06:43.000
by scientists across the world.

06:43.000 --> 06:49.000
And here's this selection of examples to keep it kind of real

06:49.000 --> 06:51.000
for what is it used for.

06:51.000 --> 06:56.000
So from Virginia Tech, they created a system to track the health of cattle.

06:56.000 --> 07:00.000
So using an experimenter that is mounted on the cow.

07:00.000 --> 07:04.000
And in that system, you automatically

07:04.000 --> 07:07.000
reanalyze with a decision tree model.

07:07.000 --> 07:10.000
The extra amount of data in determining if the cow is flying,

07:10.000 --> 07:13.000
walking, standing, grazing, or resonating.

07:13.000 --> 07:16.000
These are like the stages of food processing for the cow.

07:16.000 --> 07:19.000
And then you transmit this data over lower one.

07:19.000 --> 07:22.000
Just the not that raw data, but the activities,

07:22.000 --> 07:25.000
like the classes, what is this cow up to.

07:25.000 --> 07:29.000
And then this is used then to track, to check for abnormal behavior.

07:29.000 --> 07:32.000
So like if the cows are not eating for a long time,

07:32.000 --> 07:33.000
that's concerning.

07:33.000 --> 07:35.000
That's something you might want to check out.

07:35.000 --> 07:38.000
And doing this all on the sensor,

07:38.000 --> 07:41.000
enables it to run in their one milliwat,

07:41.000 --> 07:45.000
which is 50 times lower power consumption and sending the raw data.

07:45.000 --> 07:47.000
Which is really important in this kind of case,

07:47.000 --> 07:52.000
because basically it allows you to have 50 times the battery life.

07:52.000 --> 07:55.000
Here's a project from Samsung Research,

07:55.000 --> 07:59.000
where they estimate the breeding rate of the person wearing a

07:59.000 --> 08:03.000
this kind of earbola or device.

08:03.000 --> 08:07.000
And they use a combination of audio data as well as the

08:07.000 --> 08:10.000
extraometer data to determine how quickly you're breeding,

08:10.000 --> 08:14.000
which is very important to track for those that have respiratory health problems.

08:15.000 --> 08:19.000
And this is a practical solution that can be used in worn everyday,

08:19.000 --> 08:22.000
not just in a clinical context.

08:22.000 --> 08:26.000
And again, this is important that the power consumption is not too high,

08:26.000 --> 08:30.000
so that you can still use this for like a couple of days without having to

08:30.000 --> 08:32.000
worry about battery life.

08:32.000 --> 08:35.000
Here's a more like fun project from of mine,

08:35.000 --> 08:41.000
using an extraometer in a timer that can be attached to a toothbrush,

08:41.000 --> 08:44.000
and you automatically track how long you're actively brushing your teeth.

08:44.000 --> 08:48.000
Because you are supposed to brush your teeth for at least two minutes,

08:48.000 --> 08:51.000
every time, and most of us don't do it.

08:51.000 --> 08:54.000
So this device can help you.

08:54.000 --> 08:57.000
So this is something that a Christmas, just a fun,

08:57.000 --> 09:00.000
and one thing I'd like to highlight here is that I collected and

09:00.000 --> 09:04.000
labeled the data for this in just a couple of hours.

09:04.000 --> 09:09.000
So these kind of projects are possible also with a low amount of data,

09:09.000 --> 09:12.000
because the problem is not so complex,

09:12.000 --> 09:17.000
like a realistic type of shading that you do in the University,

09:17.000 --> 09:22.000
that can be done by just selecting like five sessions of training data

09:22.000 --> 09:24.000
with two minutes each time.

09:24.000 --> 09:28.000
So that makes it really fun to get into.

09:28.000 --> 09:30.000
Like you don't need as massive data sets,

09:30.000 --> 09:33.000
you can go directly and solve things that you're interested in

09:33.000 --> 09:37.000
around your house or home or business or whatever.

09:37.000 --> 09:40.000
So there's full example code on that available online.

09:40.000 --> 09:44.000
Here's like more of a research project of mine,

09:44.000 --> 09:49.000
like trying to really find out like how cheap can we make this system,

09:49.000 --> 09:50.000
like it's still a useful system.

09:50.000 --> 09:54.000
And I post the question, can we make a useful animal system

09:54.000 --> 09:58.000
for under $1 in total component costs?

09:58.000 --> 10:00.000
And the preliminary answer is yes,

10:00.000 --> 10:03.000
you can do either motion analysis,

10:03.000 --> 10:06.000
so use an extrometer and you need a battery,

10:06.000 --> 10:09.000
and the BLE transmission all that for under $1.

10:09.000 --> 10:13.000
So there's like the bomb is around 70 cents for that.

10:13.000 --> 10:17.000
And the model needs to fit that in the microintroller we have for that,

10:17.000 --> 10:20.000
which has 8 kilowatts of RAM and 64 kilowatts of RAM.

10:20.000 --> 10:23.000
And that's actually quite a lot from my perspective.

10:23.000 --> 10:28.000
So we can should also be able to do audio analysis with that

10:28.000 --> 10:29.000
with small neural networks.

10:29.000 --> 10:31.000
So that is also possible.

10:32.000 --> 10:33.000
So you can't do both.

10:33.000 --> 10:34.000
You can't have a,

10:34.000 --> 10:36.000
you don't have enough money for both sensors,

10:36.000 --> 10:39.000
but you can do either.

10:39.000 --> 10:44.000
So Emlearn has this C library.

10:44.000 --> 10:51.000
So that's the easiest way to deploy on the vast majority of devices,

10:51.000 --> 10:54.000
because he's the most common.

10:54.000 --> 10:59.000
And you train the process using the standard Python ML libraries.

10:59.000 --> 11:02.000
So for example, keras or psychic learn.

11:02.000 --> 11:06.000
And yeah, we support selection of different models there.

11:06.000 --> 11:11.000
Then you use the EM Learn library with this convert function to

11:11.000 --> 11:13.000
convert into an efficient model.

11:13.000 --> 11:17.000
We do a couple of simple optimizations there.

11:17.000 --> 11:22.000
And then you can use the same function to generate some C code in this case.

11:22.000 --> 11:25.000
So this is just some example code or like,

11:25.000 --> 11:27.000
I'll sample output for a neural network.

11:27.000 --> 11:32.000
And we'll look quite similar for the other kinds of models.

11:32.000 --> 11:36.000
You can then include the generated header file.

11:36.000 --> 11:42.000
And you construct your input data that you would typically read from your sensor.

11:42.000 --> 11:47.000
And it might also include some feature engineering or pre-processing there.

11:47.000 --> 11:52.000
And then you pass the data to the predicts function that we provide.

11:52.000 --> 11:59.000
And that's all you need in order to run machine learning models on these kind of devices.

11:59.000 --> 12:04.000
The alternative is to the EM Learn C library.

12:04.000 --> 12:07.000
It's used EM Learn Micro Python library.

12:07.000 --> 12:11.000
Because as I mentioned many that to machine learning,

12:11.000 --> 12:14.000
they speak primarily Python.

12:14.000 --> 12:20.000
And it's very interesting to be able to use those skills directly on these kind of devices.

12:20.000 --> 12:24.000
So Micro Python implements a subset of Python 3.

12:24.000 --> 12:30.000
It's a really nice practical way of getting into these kind of microcontroller devices.

12:30.000 --> 12:35.000
And they have a really nice feature, which is that you can load C modules at runtime.

12:35.000 --> 12:44.000
So similar as on a real computer, you can just do MIP install and a MPI native module.

12:44.000 --> 12:49.000
And then you can drop in efficient C modules that you can use from Python.

12:49.000 --> 12:52.000
And that's quite unique in this embedded space.

12:52.000 --> 12:54.000
Very useful.

12:54.000 --> 12:58.000
So we provide them a range of modules.

12:58.000 --> 12:59.000
They're standalone.

12:59.000 --> 13:01.000
So you can install just what you need.

13:01.000 --> 13:06.000
For example, infinite impulse response filters for doing digital filtering.

13:06.000 --> 13:12.000
Very common to do random forest or tree based ensemble models.

13:12.000 --> 13:15.000
They're very power efficient and small.

13:15.000 --> 13:16.000
They're very useful.

13:16.000 --> 13:18.000
And also fast for a transport.

13:18.000 --> 13:25.000
So we saw that it's useful to also provide some DSP functionality because it was a bit lacking in the Micro Python space.

13:25.000 --> 13:28.000
And you can do convolutional neural networks.

13:28.000 --> 13:35.000
So there's examples doing image classification, et cetera, available in the repository.

13:35.000 --> 13:44.000
And doing this is at least 5 to 20 times faster than, for example, generating Python code and executing that with Micro Python.

13:44.000 --> 13:55.000
So you get the convenience of using Python, but you get the speed benefits and the size benefits of the C modules.

13:55.000 --> 13:58.000
And the process of using that is very similar.

13:58.000 --> 13:59.000
You train your model.

13:59.000 --> 14:01.000
Use the convert function.

14:01.000 --> 14:07.000
In this case, we often save the model as a file that because Micro Python has a file system.

14:07.000 --> 14:13.000
So we can copy the model definition over as a loadable file.

14:13.000 --> 14:19.000
And then, yeah, when installs the library, we're just one line as well.

14:19.000 --> 14:21.000
You can create the model.

14:21.000 --> 14:23.000
You have to specify the capacity.

14:23.000 --> 14:27.000
And then we load these the model definition from the file system.

14:27.000 --> 14:29.000
And then again, you read your sensor data.

14:29.000 --> 14:33.000
And you pass the data to the model predict function.

14:33.000 --> 14:40.000
And you get the classification or regression or anomaly detection output.

14:40.000 --> 14:44.000
So that's all that we had.

14:44.000 --> 14:51.000
So in summary, machine learning's used in embedded systems to automatically analyze sensor data.

14:51.000 --> 14:53.000
That's the predominant use case.

14:53.000 --> 14:57.000
And then we can make practical applications in this space.

14:57.000 --> 14:59.000
We just, a few kilobytes of RAM.

14:59.000 --> 15:02.000
And within just a few million watts of power.

15:02.000 --> 15:07.000
Or just a few dollars in terms of build materials or hardware costs.

15:08.000 --> 15:13.000
And EM Learn is an open source project to help deploy machine learning models to microcontrollers.

15:13.000 --> 15:18.000
And you can use it either with a C library or micro Python library.

15:18.000 --> 15:20.000
So there's some resources.

15:20.000 --> 15:25.000
And you'll find in the slides, of course, a line with more information, including some other talks,

15:25.000 --> 15:34.000
and go a little bit more into depth in particular aspects of this niche and the library.

15:34.000 --> 15:37.000
So thank you so much.

15:37.000 --> 15:40.000
Thank you. Thank you very much.

15:40.000 --> 15:44.000
This, this two graphs was, this two brush.

15:44.000 --> 15:46.000
I have to buy it like this.

15:46.000 --> 15:48.000
No, you know, it opensource.

15:48.000 --> 15:51.000
So you're just download the three different files.

15:51.000 --> 15:52.000
You have three different theory yourself.

15:52.000 --> 15:53.000
Use it there.

15:53.000 --> 15:56.000
If not, you send it to someone print and you download the code.

15:56.000 --> 15:57.000
Okay, that's cool.

15:57.000 --> 15:59.000
Any questions?

15:59.000 --> 16:00.000
Yes, please.

16:00.000 --> 16:01.000
I'm shouted.

16:02.000 --> 16:03.000
There's something.

16:03.000 --> 16:07.000
How come this as a small device has some kind of a co-position?

16:07.000 --> 16:08.000
Like a vector.

16:08.000 --> 16:10.000
And can you explain on this network?

16:10.000 --> 16:11.000
Yeah.

16:11.000 --> 16:12.000
So the question.

16:12.000 --> 16:13.000
I'll repeat it.

16:13.000 --> 16:18.000
So how are there devices that have some co-processors that you can use like vector processing

16:18.000 --> 16:19.000
and stuff like that?

16:19.000 --> 16:21.000
Is it common or not?

16:21.000 --> 16:22.000
Yep.

16:22.000 --> 16:25.000
It's getting more common because this is,

16:25.000 --> 16:29.000
using doing machine learning is more and more common and relevant.

16:29.000 --> 16:30.000
But hardware takes time, right?

16:30.000 --> 16:33.000
So it's not like 99 will have it.

16:33.000 --> 16:35.000
But there are some recent instructions.

16:35.000 --> 16:37.000
So arm has like standardized,

16:37.000 --> 16:40.000
Cindy instructions for the cortex and four.

16:40.000 --> 16:41.000
That is what useful.

16:41.000 --> 16:46.000
Can you give it like a 4x benefit on like neural network type problems?

16:46.000 --> 16:48.000
And then there's helium that is coming.

16:48.000 --> 16:50.000
They have standardized it like four years ago.

16:50.000 --> 16:52.000
But devices are not really out there.

16:52.000 --> 16:55.000
But then there are on the ESP32.

16:55.000 --> 16:58.000
There's S3, which has a bunch of smaller vector in the system.

16:58.000 --> 17:00.000
Smaller vector instructions.

17:00.000 --> 17:01.000
So it's possible.

17:01.000 --> 17:07.000
It's getting possible to do like a get like a 4 to 10x type speedups.

17:07.000 --> 17:14.000
And then there are now like SC just announced a new microcontroller just before Christmas.

17:14.000 --> 17:17.000
That has like a dedicated NPU.

17:17.000 --> 17:19.000
They call it neural processing unit.

17:19.000 --> 17:23.000
So definitely this is becoming more and more common.

17:23.000 --> 17:29.000
On the high end, like if you do like powerful image classification or object detection,

17:29.000 --> 17:32.000
for example, you want to use that.

17:32.000 --> 17:35.000
But then if you're doing with extra amount of data for example,

17:35.000 --> 17:37.000
it's usually overkill.

17:37.000 --> 17:40.000
And sometimes it can be less power efficient because you have this copercer,

17:40.000 --> 17:42.000
you need to shift data in and out.

17:42.000 --> 17:44.000
So it's definitely getting there.

17:44.000 --> 17:47.000
But you can do quite a lot without them right now.

17:47.000 --> 17:49.000
So let's go.

17:49.000 --> 17:52.000
We will see how much we implement.

17:52.000 --> 17:55.000
Like those instructions.

17:55.000 --> 17:56.000
They want to get common.

17:56.000 --> 17:57.000
Yes.

17:57.000 --> 17:59.000
But right now we prefer just doing plain.

17:59.000 --> 18:01.000
See runs everywhere.

18:01.000 --> 18:03.000
Simple and usually enough.

18:03.000 --> 18:04.000
That's a question.

18:04.000 --> 18:05.000
One question from mine.

18:05.000 --> 18:10.000
Do you see any time when those microcontrolls will be arming LLM inference?

18:10.000 --> 18:11.000
LLM.

18:11.000 --> 18:12.000
Oh yes.

18:12.000 --> 18:17.000
This is there's a surprising amount of buzz about LLMs.

18:17.000 --> 18:22.000
And I don't quite understand the use cases yet.

18:22.000 --> 18:24.000
But it might be relevant for Baltics.

18:24.000 --> 18:28.000
But there's a lot of hype around it.

18:28.000 --> 18:32.000
We will see in a few years if there's like useful things or not.

18:32.000 --> 18:33.000
I don't really know.

18:33.000 --> 18:35.000
But you have definitely a lot of hype.

18:35.000 --> 18:36.000
I'd like to talk to you.

18:36.000 --> 18:37.000
I'd like to talk to you.

18:37.000 --> 18:38.000
I'd like to talk to you.

18:38.000 --> 18:40.000
I'd like to tell you you need to do more, more,

18:40.000 --> 18:43.000
like if personal trainer, I don't know.

18:43.000 --> 18:44.000
Thank you.

18:44.000 --> 18:45.000
Thank you again.

18:45.000 --> 18:46.000
Thank you.

18:46.000 --> 18:48.000
And that was the last talk.

18:48.000 --> 18:53.000
And let me pass the mic to time.