WEBVTT

025a3067-1e27-4a28-af18-1f93272d4d3c/38-0
00:00:03.579 --> 00:00:06.918
For attending the Earthquake
Science Center Weekly seminar

025a3067-1e27-4a28-af18-1f93272d4d3c/38-1
00:00:06.918 --> 00:00:08.729
series, if you are new, welcome.

025a3067-1e27-4a28-af18-1f93272d4d3c/53-0
00:00:08.779 --> 00:00:11.292
If you would like to be added to
our email distribution group,

025a3067-1e27-4a28-af18-1f93272d4d3c/53-1
00:00:11.292 --> 00:00:12.249
please send us an email.

025a3067-1e27-4a28-af18-1f93272d4d3c/99-0
00:00:12.739 --> 00:00:16.117
Seminars are recorded and mostly
all talks are posted on the USGS

025a3067-1e27-4a28-af18-1f93272d4d3c/99-1
00:00:16.117 --> 00:00:19.340
Earthquake Science Center web
page and close captioning can be

025a3067-1e27-4a28-af18-1f93272d4d3c/99-2
00:00:19.340 --> 00:00:22.615
turned on by clicking on the CC
icon in the more tab at the top

025a3067-1e27-4a28-af18-1f93272d4d3c/99-3
00:00:22.615 --> 00:00:23.229
of the page.

025a3067-1e27-4a28-af18-1f93272d4d3c/106-0
00:00:23.479 --> 00:00:25.209
Attendees, please mute your
mics.

025a3067-1e27-4a28-af18-1f93272d4d3c/157-0
00:00:25.219 --> 00:00:28.325
Turn off your cameras and until
the Q&A session at the end

025a3067-1e27-4a28-af18-1f93272d4d3c/157-1
00:00:28.325 --> 00:00:31.480
of the talk, feel free to submit
your questions via the chat at

025a3067-1e27-4a28-af18-1f93272d4d3c/157-2
00:00:31.480 --> 00:00:34.536
any time, or you can wait to
turn on your camera and ask your

025a3067-1e27-4a28-af18-1f93272d4d3c/157-3
00:00:34.536 --> 00:00:35.029
questions.

025a3067-1e27-4a28-af18-1f93272d4d3c/170-0
00:00:35.039 --> 00:00:39.569
During the Q&A session
itself, so on to announcements

025a3067-1e27-4a28-af18-1f93272d4d3c/170-1
00:00:39.569 --> 00:00:40.349
for today.

025a3067-1e27-4a28-af18-1f93272d4d3c/211-0
00:00:40.519 --> 00:00:44.159
Friday is Veterans Day, which is
a federal holiday and will be

025a3067-1e27-4a28-af18-1f93272d4d3c/211-1
00:00:44.159 --> 00:00:47.394
extremely fortunate to have a
number of veterans in the

025a3067-1e27-4a28-af18-1f93272d4d3c/211-2
00:00:47.394 --> 00:00:50.514
Science Center served and
continue it and continue to

025a3067-1e27-4a28-af18-1f93272d4d3c/211-3
00:00:50.514 --> 00:00:51.669
serve in some cases.

025a3067-1e27-4a28-af18-1f93272d4d3c/227-0
00:00:52.279 --> 00:00:55.379
So thank you to all veterans for
your service and sacrifices.

025a3067-1e27-4a28-af18-1f93272d4d3c/245-0
00:00:56.299 --> 00:00:59.977
Training completion is required
for the transit and bicycle

025a3067-1e27-4a28-af18-1f93272d4d3c/245-1
00:00:59.977 --> 00:01:03.409
subsidy program, recertification
or for new applicants.

025a3067-1e27-4a28-af18-1f93272d4d3c/256-0
00:01:03.419 --> 00:01:06.809
So completion of that training
is due by November 30th.

025a3067-1e27-4a28-af18-1f93272d4d3c/278-0
00:01:06.919 --> 00:01:10.353
For more details, see the email
from Nancy Roundtree, Esse

025a3067-1e27-4a28-af18-1f93272d4d3c/278-1
00:01:10.353 --> 00:01:13.846
coordinator, and finally it's
that time of year considering

025a3067-1e27-4a28-af18-1f93272d4d3c/278-2
00:01:13.846 --> 00:01:14.369
donating.

025a3067-1e27-4a28-af18-1f93272d4d3c/292-0
00:01:14.379 --> 00:01:17.150
Consider donating to your
favorite charity using the

025a3067-1e27-4a28-af18-1f93272d4d3c/292-1
00:01:17.150 --> 00:01:18.509
combined federal campaign.

025a3067-1e27-4a28-af18-1f93272d4d3c/304-0
00:01:18.699 --> 00:01:21.349
It's open until January 15th,
2024.

025a3067-1e27-4a28-af18-1f93272d4d3c/335-0
00:01:21.459 --> 00:01:25.070
The CFC gives federal employees
a chance to come together to

025a3067-1e27-4a28-af18-1f93272d4d3c/335-1
00:01:25.070 --> 00:01:28.444
help charities across the
country see your inbox for the

025a3067-1e27-4a28-af18-1f93272d4d3c/335-2
00:01:28.444 --> 00:01:30.929
email from the secretary of the
Interior.

025a3067-1e27-4a28-af18-1f93272d4d3c/349-0
00:01:31.739 --> 00:01:34.660
With that, I will hand it over
to Tim, who will introduce our

025a3067-1e27-4a28-af18-1f93272d4d3c/349-1
00:01:34.660 --> 00:01:35.319
speaker today.

025a3067-1e27-4a28-af18-1f93272d4d3c/365-0
00:01:38.159 --> 00:01:41.709
And so happy to have army
Novoselov here from Stanford.

025a3067-1e27-4a28-af18-1f93272d4d3c/392-0
00:01:42.659 --> 00:01:46.669
He's currently a postdoc working
with Greg Beroza, so previously

025a3067-1e27-4a28-af18-1f93272d4d3c/392-1
00:01:46.669 --> 00:01:50.369
that he is a BS in Geoscience
from Moscow State University.

025a3067-1e27-4a28-af18-1f93272d4d3c/414-0
00:01:50.939 --> 00:01:53.549
He didn't completed a Masters
degree in machine learning and

025a3067-1e27-4a28-af18-1f93272d4d3c/414-1
00:01:53.549 --> 00:01:55.859
geoscience at the Norwegian
University of Science and

025a3067-1e27-4a28-af18-1f93272d4d3c/414-2
00:01:55.859 --> 00:01:56.329
Technology.

025a3067-1e27-4a28-af18-1f93272d4d3c/447-0
00:01:57.499 --> 00:02:00.935
I didn't get a PhD at the
University of Vienna working on

025a3067-1e27-4a28-af18-1f93272d4d3c/447-1
00:02:00.935 --> 00:02:04.489
kind of time Series analysis
supplied to seismic data using

025a3067-1e27-4a28-af18-1f93272d4d3c/447-2
00:02:04.489 --> 00:02:06.029
machine and deep learning.

025a3067-1e27-4a28-af18-1f93272d4d3c/459-0
00:02:06.569 --> 00:02:09.649
Umm so I meant admin multiple
conference.

025a3067-1e27-4a28-af18-1f93272d4d3c/476-0
00:02:11.009 --> 00:02:13.816
Post your halls in the deep
learning section and so I'm

025a3067-1e27-4a28-af18-1f93272d4d3c/476-1
00:02:13.816 --> 00:02:15.319
really happy to have him here.

025a3067-1e27-4a28-af18-1f93272d4d3c/504-0
00:02:15.469 --> 00:02:18.832
I thought this was super cool to
have him looking at face picking

025a3067-1e27-4a28-af18-1f93272d4d3c/504-1
00:02:18.832 --> 00:02:21.685
in it in a different way to
generalized way, so give it

025a3067-1e27-4a28-af18-1f93272d4d3c/504-2
00:02:21.685 --> 00:02:21.939
time.

025a3067-1e27-4a28-af18-1f93272d4d3c/515-0
00:02:25.939 --> 00:02:27.269
Thank you, team for
introduction.

025a3067-1e27-4a28-af18-1f93272d4d3c/554-0
00:02:27.279 --> 00:02:31.542
I'm super happy to be here today
and so let's let's go through

025a3067-1e27-4a28-af18-1f93272d4d3c/554-1
00:02:31.542 --> 00:02:35.398
beyond face speaking face
hunters generalizable approach

025a3067-1e27-4a28-af18-1f93272d4d3c/554-2
00:02:35.398 --> 00:02:39.389
to seismic signal analysis using
deep learning regression.

025a3067-1e27-4a28-af18-1f93272d4d3c/557-0
00:02:39.779 --> 00:02:45.149
And let's start by talking about
face speaking.

025a3067-1e27-4a28-af18-1f93272d4d3c/590-0
00:02:45.219 --> 00:02:51.932
You know, up until recently the
way how people picked and onset

025a3067-1e27-4a28-af18-1f93272d4d3c/590-1
00:02:51.932 --> 00:02:58.329
arrivals inside the data was
either manual or by using SDLT.

025a3067-1e27-4a28-af18-1f93272d4d3c/651-0
00:02:58.799 --> 00:03:02.925
SLA is when you have a short
term window and a short term

025a3067-1e27-4a28-af18-1f93272d4d3c/651-1
00:03:02.925 --> 00:03:07.050
average and the long term
average and you try to find the

025a3067-1e27-4a28-af18-1f93272d4d3c/651-2
00:03:07.050 --> 00:03:11.602
ratio between the two and then
when the ratio exceeds assertion

025a3067-1e27-4a28-af18-1f93272d4d3c/651-3
00:03:11.602 --> 00:03:16.083
threshold called trigger you say
well OK, this is the onset of

025a3067-1e27-4a28-af18-1f93272d4d3c/651-4
00:03:16.083 --> 00:03:17.149
the face right?

025a3067-1e27-4a28-af18-1f93272d4d3c/656-0
00:03:17.239 --> 00:03:18.439
And that works.

025a3067-1e27-4a28-af18-1f93272d4d3c/661-0
00:03:18.539 --> 00:03:19.569
That works pretty well.

025a3067-1e27-4a28-af18-1f93272d4d3c/667-0
00:03:19.609 --> 00:03:21.789
That works surprisingly well.

025a3067-1e27-4a28-af18-1f93272d4d3c/703-0
00:03:22.359 --> 00:03:26.730
Uh, and if you, if you say,
let's say you have a an array of

025a3067-1e27-4a28-af18-1f93272d4d3c/703-1
00:03:26.730 --> 00:03:31.030
seismic stations, you picked all
of this arrivals, then you

025a3067-1e27-4a28-af18-1f93272d4d3c/703-2
00:03:31.030 --> 00:03:32.749
located the earthquakes.

025a3067-1e27-4a28-af18-1f93272d4d3c/712-0
00:03:32.839 --> 00:03:34.769
This is an example of what you
would have.

025a3067-1e27-4a28-af18-1f93272d4d3c/730-0
00:03:36.369 --> 00:03:40.072
This works right and it was
working for the whole history of

025a3067-1e27-4a28-af18-1f93272d4d3c/730-1
00:03:40.072 --> 00:03:40.739
seismology.

025a3067-1e27-4a28-af18-1f93272d4d3c/762-0
00:03:40.749 --> 00:03:45.933
Ever since we figured out that
we need to do face speaking, the

025a3067-1e27-4a28-af18-1f93272d4d3c/762-1
00:03:45.933 --> 00:03:50.955
main problem with that is that
this SLT massively gives you a

025a3067-1e27-4a28-af18-1f93272d4d3c/762-2
00:03:50.955 --> 00:03:53.709
massive amount of post
positives.

025a3067-1e27-4a28-af18-1f93272d4d3c/787-0
00:03:54.179 --> 00:03:57.541
You would, you know, have a have
a trigger it where it shouldn't

025a3067-1e27-4a28-af18-1f93272d4d3c/787-1
00:03:57.541 --> 00:03:59.609
be and on noise and so on and so
forth.

025a3067-1e27-4a28-af18-1f93272d4d3c/829-0
00:03:59.649 --> 00:04:03.635
So also when if you have phases
that are buried in the noise,

025a3067-1e27-4a28-af18-1f93272d4d3c/829-1
00:04:03.635 --> 00:04:07.620
you can't really detect them
with the LDA unless you do a lot

025a3067-1e27-4a28-af18-1f93272d4d3c/829-2
00:04:07.620 --> 00:04:08.969
of manual processing.

025a3067-1e27-4a28-af18-1f93272d4d3c/857-0
00:04:08.979 --> 00:04:12.913
So the deal is that we can have
nice face speaking, but it

025a3067-1e27-4a28-af18-1f93272d4d3c/857-1
00:04:12.913 --> 00:04:15.779
requires a lot of manual labor
to do that.

025a3067-1e27-4a28-af18-1f93272d4d3c/901-0
00:04:15.949 --> 00:04:20.449
Then, in 2017, when new wave of
tools emerged called the deep

025a3067-1e27-4a28-af18-1f93272d4d3c/901-1
00:04:20.449 --> 00:04:24.804
learning right and the first
thing that now is the de facto

025a3067-1e27-4a28-af18-1f93272d4d3c/901-2
00:04:24.804 --> 00:04:29.158
standard is phase net, which
originated from Greg Burrows's

025a3067-1e27-4a28-af18-1f93272d4d3c/901-3
00:04:29.158 --> 00:04:30.319
lab at Stanford.

025a3067-1e27-4a28-af18-1f93272d4d3c/903-0
00:04:30.329 --> 00:04:31.539
'S lab at Stanford.

025a3067-1e27-4a28-af18-1f93272d4d3c/928-0
00:04:31.769 --> 00:04:37.136
So the way the way it works, the
phase net predicts a

025a3067-1e27-4a28-af18-1f93272d4d3c/928-1
00:04:37.136 --> 00:04:42.999
distribution of probability
where in a sample the peak is.

025a3067-1e27-4a28-af18-1f93272d4d3c/944-0
00:04:43.269 --> 00:04:46.729
Let me explain it, I will
explain it in in a second.

025a3067-1e27-4a28-af18-1f93272d4d3c/1010-0
00:04:46.739 --> 00:04:50.846
If you do that, like if you just
apply a neural network face

025a3067-1e27-4a28-af18-1f93272d4d3c/1010-1
00:04:50.846 --> 00:04:55.020
speaker, you have a much denser
detection like you can locate

025a3067-1e27-4a28-af18-1f93272d4d3c/1010-2
00:04:55.020 --> 00:04:58.588
much more events and it's
marvelous right as you all

025a3067-1e27-4a28-af18-1f93272d4d3c/1010-3
00:04:58.588 --> 00:05:02.493
probably know and the way it
works is essentially you are

025a3067-1e27-4a28-af18-1f93272d4d3c/1010-4
00:05:02.493 --> 00:05:06.465
classifying like you have a
waveform. Waveform is a number

025a3067-1e27-4a28-af18-1f93272d4d3c/1010-5
00:05:06.465 --> 00:05:07.609
of samples right?

025a3067-1e27-4a28-af18-1f93272d4d3c/1068-0
00:05:07.619 --> 00:05:11.535
Like you have, let's say 3000
samples in your waveform and

025a3067-1e27-4a28-af18-1f93272d4d3c/1068-1
00:05:11.535 --> 00:05:15.783
then with the phase that you're
classifying each of the samples

025a3067-1e27-4a28-af18-1f93272d4d3c/1052-0
00:05:14.449 --> 00:05:15.219
Oh, look at you.

025a3067-1e27-4a28-af18-1f93272d4d3c/1068-2
00:05:15.783 --> 00:05:20.098
whether it belongs to the class,
whether it belongs to the class

025a3067-1e27-4a28-af18-1f93272d4d3c/1066-0
00:05:19.479 --> 00:05:19.879
Thank you.

025a3067-1e27-4a28-af18-1f93272d4d3c/1070-0
00:05:19.889 --> 00:05:21.089
Might need a bath.

025a3067-1e27-4a28-af18-1f93272d4d3c/1068-3
00:05:20.098 --> 00:05:20.429
of P.

025a3067-1e27-4a28-af18-1f93272d4d3c/1085-0
00:05:22.119 --> 00:05:27.729
I this is not believe someone is
not muted and alright.

025a3067-1e27-4a28-af18-1f93272d4d3c/1146-0
00:05:27.739 --> 00:05:32.003
So the again, you have 3000
samples and you classify each of

025a3067-1e27-4a28-af18-1f93272d4d3c/1146-1
00:05:32.003 --> 00:05:36.197
them whether it belongs to the
class P or it belongs to the

025a3067-1e27-4a28-af18-1f93272d4d3c/1146-2
00:05:36.197 --> 00:05:40.391
class South or it belongs to
noise and then you try to like

025a3067-1e27-4a28-af18-1f93272d4d3c/1146-3
00:05:40.391 --> 00:05:44.586
accumulate it in the way and
then you have like a sort of a

025a3067-1e27-4a28-af18-1f93272d4d3c/1146-4
00:05:44.586 --> 00:05:48.849
distribution where the peak
should be just like that, right.

025a3067-1e27-4a28-af18-1f93272d4d3c/1162-0
00:05:48.859 --> 00:05:52.469
So you have like just classify
under samples from zero to 1.

025a3067-1e27-4a28-af18-1f93272d4d3c/1175-0
00:05:53.579 --> 00:05:57.179
The problem with this approach
is that, umm, it works.

025a3067-1e27-4a28-af18-1f93272d4d3c/1196-0
00:05:57.369 --> 00:06:02.432
It's great, but you have what
first you you train the whole

025a3067-1e27-4a28-af18-1f93272d4d3c/1196-1
00:06:02.432 --> 00:06:04.879
thing on truncated Gaussians.

025a3067-1e27-4a28-af18-1f93272d4d3c/1241-0
00:06:04.889 --> 00:06:08.209
Basically, the way the train
phase net is, you take you you

025a3067-1e27-4a28-af18-1f93272d4d3c/1241-1
00:06:08.209 --> 00:06:11.638
know where you sample is you
generate the Gaussian around the

025a3067-1e27-4a28-af18-1f93272d4d3c/1241-2
00:06:11.638 --> 00:06:14.902
sample and you say, well, we
have this awesome probability

025a3067-1e27-4a28-af18-1f93272d4d3c/1241-3
00:06:14.902 --> 00:06:17.889
distribution that has nothing to
do with probability.

025a3067-1e27-4a28-af18-1f93272d4d3c/1261-0
00:06:17.899 --> 00:06:21.722
It's just a Gaussian distributed
in space and now we train and it

025a3067-1e27-4a28-af18-1f93272d4d3c/1261-1
00:06:21.722 --> 00:06:22.069
works.

025a3067-1e27-4a28-af18-1f93272d4d3c/1267-0
00:06:22.079 --> 00:06:23.969
It works marvelous, right?

025a3067-1e27-4a28-af18-1f93272d4d3c/1303-0
00:06:24.019 --> 00:06:28.040
It's surprising how it works and
the other thing is that when you

025a3067-1e27-4a28-af18-1f93272d4d3c/1303-1
00:06:28.040 --> 00:06:31.877
classify samples, you can only
achieve the accuracy of samples

025a3067-1e27-4a28-af18-1f93272d4d3c/1303-2
00:06:31.877 --> 00:06:34.679
like you cannot go beyond your
sampling rate.

025a3067-1e27-4a28-af18-1f93272d4d3c/1356-0
00:06:34.689 --> 00:06:38.112
You cannot go beyond your
sampling resolution, so there

025a3067-1e27-4a28-af18-1f93272d4d3c/1356-1
00:06:38.112 --> 00:06:41.718
might be a better way of doing
that and or not necessarily

025a3067-1e27-4a28-af18-1f93272d4d3c/1356-2
00:06:41.718 --> 00:06:45.507
better, but a different way of
doing that which allows you to

025a3067-1e27-4a28-af18-1f93272d4d3c/1356-3
00:06:45.507 --> 00:06:48.807
have at least theoretical
subsample accuracy which is

025a3067-1e27-4a28-af18-1f93272d4d3c/1356-4
00:06:48.807 --> 00:06:49.479
regression.

025a3067-1e27-4a28-af18-1f93272d4d3c/1376-0
00:06:49.809 --> 00:06:54.901
And when we are trying to have
our own set times, we want to

025a3067-1e27-4a28-af18-1f93272d4d3c/1376-1
00:06:54.901 --> 00:06:56.069
have a number.

025a3067-1e27-4a28-af18-1f93272d4d3c/1390-0
00:06:56.159 --> 00:06:59.421
We don't really wanna have this
whole distribution Gaussian

025a3067-1e27-4a28-af18-1f93272d4d3c/1390-1
00:06:59.421 --> 00:07:00.399
probability thing.

025a3067-1e27-4a28-af18-1f93272d4d3c/1408-0
00:07:00.409 --> 00:07:04.309
We just wanna have a number like
the phase arrives at 5 seconds.

025a3067-1e27-4a28-af18-1f93272d4d3c/1415-0
00:07:04.319 --> 00:07:06.529
The S phase arrives at 7
seconds.

025a3067-1e27-4a28-af18-1f93272d4d3c/1420-0
00:07:06.599 --> 00:07:07.479
That's what we want.

025a3067-1e27-4a28-af18-1f93272d4d3c/1428-0
00:07:08.449 --> 00:07:11.519
I'm so here comes the face
Hunter.

025a3067-1e27-4a28-af18-1f93272d4d3c/1443-0
00:07:11.829 --> 00:07:15.629
So the way face Hunter works
will be explained in a second.

025a3067-1e27-4a28-af18-1f93272d4d3c/1469-0
00:07:15.749 --> 00:07:19.036
So face Hunter is a neural
network and the same kind of

025a3067-1e27-4a28-af18-1f93272d4d3c/1469-1
00:07:19.036 --> 00:07:21.559
sense as phase net as an error
on network.

025a3067-1e27-4a28-af18-1f93272d4d3c/1482-0
00:07:21.669 --> 00:07:24.219
We have an input which is A3
component waveform.

025a3067-1e27-4a28-af18-1f93272d4d3c/1499-0
00:07:25.169 --> 00:07:28.093
Then we have a feature
extraction module which is a

025a3067-1e27-4a28-af18-1f93272d4d3c/1499-1
00:07:28.093 --> 00:07:29.779
component of a neural network.

025a3067-1e27-4a28-af18-1f93272d4d3c/1521-0
00:07:29.789 --> 00:07:32.909
Think of it as just a like a big
chunk of a neural network that

025a3067-1e27-4a28-af18-1f93272d4d3c/1521-1
00:07:32.909 --> 00:07:34.469
is composed of different layers.

025a3067-1e27-4a28-af18-1f93272d4d3c/1561-0
00:07:35.639 --> 00:07:40.430
What's happening here is that we
are transforming the input

025a3067-1e27-4a28-af18-1f93272d4d3c/1561-1
00:07:40.430 --> 00:07:44.581
signal to some higher
dimensional representation of

025a3067-1e27-4a28-af18-1f93272d4d3c/1561-2
00:07:44.581 --> 00:07:49.291
the signal, kind of similar in
the spirit to what we do is

025a3067-1e27-4a28-af18-1f93272d4d3c/1561-3
00:07:49.291 --> 00:07:50.169
short STFT.

025a3067-1e27-4a28-af18-1f93272d4d3c/1577-0
00:07:51.199 --> 00:07:54.249
Uh, but different in, you know,
the way it's done.

025a3067-1e27-4a28-af18-1f93272d4d3c/1592-0
00:07:54.319 --> 00:07:57.259
It's like a transformation of a
signal to some higher

025a3067-1e27-4a28-af18-1f93272d4d3c/1592-1
00:07:57.259 --> 00:07:58.729
dimensional representation.

025a3067-1e27-4a28-af18-1f93272d4d3c/1595-0
00:07:59.339 --> 00:07:59.769
Nice.

025a3067-1e27-4a28-af18-1f93272d4d3c/1628-0
00:07:59.779 --> 00:08:03.176
And we do that by running these
signal through a series of

025a3067-1e27-4a28-af18-1f93272d4d3c/1628-1
00:08:03.176 --> 00:08:06.629
layers in the neural network,
which are not that important.

025a3067-1e27-4a28-af18-1f93272d4d3c/1659-0
00:08:08.199 --> 00:08:12.424
At the end we have this fancy
higher dimensional

025a3067-1e27-4a28-af18-1f93272d4d3c/1659-1
00:08:12.424 --> 00:08:18.114
representation of of our signal,
and now we want to predict where

025a3067-1e27-4a28-af18-1f93272d4d3c/1659-2
00:08:18.114 --> 00:08:20.269
the PNS phase phases are.

025a3067-1e27-4a28-af18-1f93272d4d3c/1670-0
00:08:20.339 --> 00:08:22.409
So for that we use a regression
module.

025a3067-1e27-4a28-af18-1f93272d4d3c/1693-0
00:08:22.559 --> 00:08:26.094
Regression module is the
simplest piece of neural network

025a3067-1e27-4a28-af18-1f93272d4d3c/1693-1
00:08:26.094 --> 00:08:28.409
available in in your of the
packages.

025a3067-1e27-4a28-af18-1f93272d4d3c/1718-0
00:08:28.419 --> 00:08:31.778
It's just a linear layer, so we
have an input which is a vector

025a3067-1e27-4a28-af18-1f93272d4d3c/1718-1
00:08:31.778 --> 00:08:33.929
and we have an output which is a
vector.

025a3067-1e27-4a28-af18-1f93272d4d3c/1737-0
00:08:34.019 --> 00:08:37.569
So if you're just predicting P
phase, you're predicting a

025a3067-1e27-4a28-af18-1f93272d4d3c/1737-1
00:08:37.569 --> 00:08:38.609
vector of size 1.

025a3067-1e27-4a28-af18-1f93272d4d3c/1794-0
00:08:38.679 --> 00:08:42.792
So you had a vector of features
of size 1000, let's say, and

025a3067-1e27-4a28-af18-1f93272d4d3c/1794-1
00:08:42.792 --> 00:08:46.836
you're predicting just one, and
you know the output of this

025a3067-1e27-4a28-af18-1f93272d4d3c/1794-2
00:08:46.836 --> 00:08:51.218
prediction is a number from zero
to five representing percent in

025a3067-1e27-4a28-af18-1f93272d4d3c/1794-3
00:08:51.218 --> 00:08:51.959
the sample.

025a3067-1e27-4a28-af18-1f93272d4d3c/1806-0
00:08:52.299 --> 00:08:55.631
So, you know zero will be 0
seconds and one would be like 30

025a3067-1e27-4a28-af18-1f93272d4d3c/1806-1
00:08:55.631 --> 00:08:57.979
seconds if the whole sample is
30 seconds.

025a3067-1e27-4a28-af18-1f93272d4d3c/1830-0
00:08:58.069 --> 00:09:02.153
So 050.5 will be 15 seconds and
this is great because you don't

025a3067-1e27-4a28-af18-1f93272d4d3c/1830-1
00:09:02.153 --> 00:09:05.279
really need to know how many
samples there were.

025a3067-1e27-4a28-af18-1f93272d4d3c/1838-0
00:09:05.289 --> 00:09:07.809
You just always have a window
and it's fine.

025a3067-1e27-4a28-af18-1f93272d4d3c/1872-0
00:09:08.739 --> 00:09:13.452
And if you wanted to predict 2
phases, you can just have a

025a3067-1e27-4a28-af18-1f93272d4d3c/1872-1
00:09:13.452 --> 00:09:18.323
linear layer that transforms
from 1000 to 2 to the vector of

025a3067-1e27-4a28-af18-1f93272d4d3c/1872-2
00:09:18.323 --> 00:09:20.719
size 2 and so on and so forth.

025a3067-1e27-4a28-af18-1f93272d4d3c/1915-0
00:09:20.729 --> 00:09:23.870
So if you wanna have an outputs
in an error network, you just

025a3067-1e27-4a28-af18-1f93272d4d3c/1915-1
00:09:23.870 --> 00:09:27.061
say, set a single parameter in
the whole error network and now

025a3067-1e27-4a28-af18-1f93272d4d3c/1915-2
00:09:27.061 --> 00:09:29.289
you are able to output multiple
parameters.

025a3067-1e27-4a28-af18-1f93272d4d3c/1951-0
00:09:30.979 --> 00:09:35.033
If you would compare to a phase
net, we would see that they both

025a3067-1e27-4a28-af18-1f93272d4d3c/1951-1
00:09:35.033 --> 00:09:38.275
neural networks and the
difference is that well the

025a3067-1e27-4a28-af18-1f93272d4d3c/1951-2
00:09:38.275 --> 00:09:40.769
architectures are different
completely.

025a3067-1e27-4a28-af18-1f93272d4d3c/1979-0
00:09:40.779 --> 00:09:45.454
But the difference for the user
is that the phase net generates

025a3067-1e27-4a28-af18-1f93272d4d3c/1979-1
00:09:45.454 --> 00:09:49.251
an array of samples with
probability of each sample

025a3067-1e27-4a28-af18-1f93272d4d3c/1979-2
00:09:49.251 --> 00:09:51.369
belonging to a certain class.

025a3067-1e27-4a28-af18-1f93272d4d3c/2019-0
00:09:51.419 --> 00:09:56.199
So you have like 3 arrays, let's
say if the waveforms were like

025a3067-1e27-4a28-af18-1f93272d4d3c/2019-1
00:09:56.199 --> 00:10:00.978
3000 samples, you have 3 arrays
by 3000 samples and face Hunter

025a3067-1e27-4a28-af18-1f93272d4d3c/2019-2
00:10:00.978 --> 00:10:04.039
just predicts numbers, just two
numbers.

025a3067-1e27-4a28-af18-1f93272d4d3c/2027-0
00:10:04.329 --> 00:10:06.409
As the output, right?

025a3067-1e27-4a28-af18-1f93272d4d3c/2031-0
00:10:07.729 --> 00:10:08.369
Umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/2124-0
00:10:09.009 --> 00:10:12.476
As I mentioned before, in the
phase net, the probability that

025a3067-1e27-4a28-af18-1f93272d4d3c/2124-1
00:10:12.476 --> 00:10:15.718
you're predicting it's not a
real probability like it has

025a3067-1e27-4a28-af18-1f93272d4d3c/2124-2
00:10:15.718 --> 00:10:19.240
nothing to do with probabilities
as you know as they are, it's

025a3067-1e27-4a28-af18-1f93272d4d3c/2124-3
00:10:19.240 --> 00:10:22.874
just the like you say that well,
I mean it's a model probability

025a3067-1e27-4a28-af18-1f93272d4d3c/2124-4
00:10:22.874 --> 00:10:26.396
in the sense, right, you have a
Gaussian and you have a center

025a3067-1e27-4a28-af18-1f93272d4d3c/2124-5
00:10:26.396 --> 00:10:29.918
of this Gaussian representing
the best peak and then say well,

025a3067-1e27-4a28-af18-1f93272d4d3c/2124-6
00:10:29.918 --> 00:10:33.273
I mean I have this 0.5 signal
distribution Gaussian or like

025a3067-1e27-4a28-af18-1f93272d4d3c/2124-7
00:10:33.273 --> 00:10:35.229
you can change it during
training.

025a3067-1e27-4a28-af18-1f93272d4d3c/2146-0
00:10:36.039 --> 00:10:41.351
But what if we what if we are
able to predict an actual

025a3067-1e27-4a28-af18-1f93272d4d3c/2146-1
00:10:41.351 --> 00:10:42.489
probability?

025a3067-1e27-4a28-af18-1f93272d4d3c/2159-0
00:10:43.219 --> 00:10:46.099
There will be, you know,
interpreted in some weird way.

025a3067-1e27-4a28-af18-1f93272d4d3c/2172-0
00:10:47.009 --> 00:10:51.159
And the way we could do that is
by estimating an uncertainty.

025a3067-1e27-4a28-af18-1f93272d4d3c/2190-0
00:10:52.259 --> 00:10:55.679
So you know, we have a in
general, we have a literary

025a3067-1e27-4a28-af18-1f93272d4d3c/2190-1
00:10:55.679 --> 00:10:56.439
uncertainty.

025a3067-1e27-4a28-af18-1f93272d4d3c/2253-0
00:10:56.449 --> 00:11:00.085
We have a postman uncertainty
and in neural networks it's very

025a3067-1e27-4a28-af18-1f93272d4d3c/2253-1
00:11:00.085 --> 00:11:03.894
difficult to say, you know which
one is which and how to quantify

025a3067-1e27-4a28-af18-1f93272d4d3c/2253-2
00:11:03.894 --> 00:11:07.530
them, but together they combine
the predictive uncertainty and

025a3067-1e27-4a28-af18-1f93272d4d3c/2253-3
00:11:07.530 --> 00:11:11.166
which can also be interpreted as
models confidence which we're

025a3067-1e27-4a28-af18-1f93272d4d3c/2253-4
00:11:11.166 --> 00:11:13.359
going to talk about in a second,
umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/2269-0
00:11:13.569 --> 00:11:16.532
So all neural networks are
wrong, but neural networks that

025a3067-1e27-4a28-af18-1f93272d4d3c/2269-1
00:11:16.532 --> 00:11:18.339
know that they're wrong are
useful.

025a3067-1e27-4a28-af18-1f93272d4d3c/2288-0
00:11:20.709 --> 00:11:25.554
And the way you can do this
predictive uncertainty is is

025a3067-1e27-4a28-af18-1f93272d4d3c/2288-1
00:11:25.554 --> 00:11:26.659
this follows.

025a3067-1e27-4a28-af18-1f93272d4d3c/2292-0
00:11:26.789 --> 00:11:27.779
So you have.

025a3067-1e27-4a28-af18-1f93272d4d3c/2320-0
00:11:28.329 --> 00:11:31.574
Let's say you have a neural
network and the one of the

025a3067-1e27-4a28-af18-1f93272d4d3c/2320-1
00:11:31.574 --> 00:11:35.114
techniques people are using to
train a neural network is to

025a3067-1e27-4a28-af18-1f93272d4d3c/2320-2
00:11:35.114 --> 00:11:35.999
have a dropout.

025a3067-1e27-4a28-af18-1f93272d4d3c/2347-0
00:11:36.129 --> 00:11:40.931
So what dropout is is that you
have your neurons, and then you

025a3067-1e27-4a28-af18-1f93272d4d3c/2347-1
00:11:40.931 --> 00:11:42.989
suddenly keel some of them.

025a3067-1e27-4a28-af18-1f93272d4d3c/2394-0
00:11:42.999 --> 00:11:46.370
Just you know your input just
doesn't go through this neuron

025a3067-1e27-4a28-af18-1f93272d4d3c/2394-1
00:11:46.370 --> 00:11:49.961
in particular, and let's say if
you have a neural network of 100

025a3067-1e27-4a28-af18-1f93272d4d3c/2394-2
00:11:49.961 --> 00:11:53.221
theorems, you just randomly
killing and every time it runs

025a3067-1e27-4a28-af18-1f93272d4d3c/2394-3
00:11:53.221 --> 00:11:54.989
cycle you just kill and you run.

025a3067-1e27-4a28-af18-1f93272d4d3c/2425-0
00:11:55.079 --> 00:11:58.707
Then the inference time it is
recommended to turn it off

025a3067-1e27-4a28-af18-1f93272d4d3c/2425-1
00:11:58.707 --> 00:12:02.335
because you want to use all of
your neurons that now are

025a3067-1e27-4a28-af18-1f93272d4d3c/2425-2
00:12:02.335 --> 00:12:05.389
trained in like this interesting
regulated way.

025a3067-1e27-4a28-af18-1f93272d4d3c/2475-0
00:12:05.879 --> 00:12:09.564
But what if you just keep this
dropout thing on and then just

025a3067-1e27-4a28-af18-1f93272d4d3c/2475-1
00:12:09.564 --> 00:12:13.485
repeat the inference so you have
a single input which is like you

025a3067-1e27-4a28-af18-1f93272d4d3c/2475-2
00:12:13.485 --> 00:12:17.170
have the same input and you run
it multiple times through the

025a3067-1e27-4a28-af18-1f93272d4d3c/2475-3
00:12:17.170 --> 00:12:20.259
same network, which different
neurons being killed.

025a3067-1e27-4a28-af18-1f93272d4d3c/2491-0
00:12:20.329 --> 00:12:22.702
So every time you do that,
you're going to have slightly

025a3067-1e27-4a28-af18-1f93272d4d3c/2491-1
00:12:22.702 --> 00:12:23.409
different answer.

025a3067-1e27-4a28-af18-1f93272d4d3c/2507-0
00:12:24.259 --> 00:12:27.436
Then you can aggregate those
answers and estimate how much

025a3067-1e27-4a28-af18-1f93272d4d3c/2507-1
00:12:27.436 --> 00:12:28.459
different they are.

025a3067-1e27-4a28-af18-1f93272d4d3c/2566-0
00:12:30.289 --> 00:12:35.517
So in this example we have 0.5,
I mean 10.50 point 98102 and

025a3067-1e27-4a28-af18-1f93272d4d3c/2566-1
00:12:35.517 --> 00:12:40.574
altogether they combine a like a
super cool answer with an

025a3067-1e27-4a28-af18-1f93272d4d3c/2566-2
00:12:40.574 --> 00:12:45.373
uncertainty to this answer,
variability in a sense, the

025a3067-1e27-4a28-af18-1f93272d4d3c/2566-3
00:12:45.373 --> 00:12:49.229
other way of doing that is batch
assembling.

025a3067-1e27-4a28-af18-1f93272d4d3c/2591-0
00:12:50.269 --> 00:12:54.210
So in in this you trains
slightly different versions of

025a3067-1e27-4a28-af18-1f93272d4d3c/2591-1
00:12:54.210 --> 00:12:55.899
the same narrow network.

025a3067-1e27-4a28-af18-1f93272d4d3c/2626-0
00:12:56.129 --> 00:12:59.867
Let's say you use different
seeds like random seed for the

025a3067-1e27-4a28-af18-1f93272d4d3c/2626-1
00:12:59.867 --> 00:13:03.795
training and when you run your
inputs through those different

025a3067-1e27-4a28-af18-1f93272d4d3c/2626-2
00:13:03.795 --> 00:13:07.279
neural networks, you have a
slightly different answer.

025a3067-1e27-4a28-af18-1f93272d4d3c/2646-0
00:13:07.469 --> 00:13:11.436
Again, you can estimate how
differently trained neural

025a3067-1e27-4a28-af18-1f93272d4d3c/2646-1
00:13:11.436 --> 00:13:15.547
networks are, giving you
different answers to that great

025a3067-1e27-4a28-af18-1f93272d4d3c/2646-2
00:13:15.547 --> 00:13:15.979
right.

025a3067-1e27-4a28-af18-1f93272d4d3c/2664-0
00:13:17.519 --> 00:13:22.140
There is a a a more interesting
way of doing that called the

025a3067-1e27-4a28-af18-1f93272d4d3c/2664-1
00:13:22.140 --> 00:13:23.049
mask sample.

025a3067-1e27-4a28-af18-1f93272d4d3c/2679-0
00:13:23.419 --> 00:13:27.709
So a mask sample is a layer you
embed in your layer letter.

025a3067-1e27-4a28-af18-1f93272d4d3c/2706-0
00:13:27.719 --> 00:13:31.476
So you when you design A neural
network, you just put those

025a3067-1e27-4a28-af18-1f93272d4d3c/2706-1
00:13:31.476 --> 00:13:34.669
layers inside and then it works
like a Lego tower.

025a3067-1e27-4a28-af18-1f93272d4d3c/2743-0
00:13:35.519 --> 00:13:39.076
If you think of it, if you think
of a narrow network as a tower

025a3067-1e27-4a28-af18-1f93272d4d3c/2743-1
00:13:39.076 --> 00:13:42.466
of Lego blocks, let's say all of
the Lego blocks are colored

025a3067-1e27-4a28-af18-1f93272d4d3c/2743-2
00:13:42.466 --> 00:13:42.799
green.

025a3067-1e27-4a28-af18-1f93272d4d3c/2774-0
00:13:43.869 --> 00:13:49.442
Now if we insert a layer of mask
sample, let let this layer to be

025a3067-1e27-4a28-af18-1f93272d4d3c/2774-1
00:13:49.442 --> 00:13:53.579
colored yellow and and now if we
would have our.

025a3067-1e27-4a28-af18-1f93272d4d3c/2789-0
00:13:54.289 --> 00:13:57.724
This is how this is how it looks
like with the layers of you

025a3067-1e27-4a28-af18-1f93272d4d3c/2789-1
00:13:57.724 --> 00:13:58.399
know, Legos.

025a3067-1e27-4a28-af18-1f93272d4d3c/2796-0
00:13:58.489 --> 00:13:59.479
But I'm talking about.

025a3067-1e27-4a28-af18-1f93272d4d3c/2834-0
00:13:59.549 --> 00:14:03.331
So now we have our input that is
flowing through the network, but

025a3067-1e27-4a28-af18-1f93272d4d3c/2834-1
00:14:03.331 --> 00:14:06.711
because of the mask samples this
inputs is flowing through

025a3067-1e27-4a28-af18-1f93272d4d3c/2834-2
00:14:06.711 --> 00:14:08.429
different paths in the narrow.

025a3067-1e27-4a28-af18-1f93272d4d3c/2839-0
00:14:08.439 --> 00:14:08.869
That's right.

025a3067-1e27-4a28-af18-1f93272d4d3c/2847-0
00:14:09.449 --> 00:14:10.979
We have the same narrow network.

025a3067-1e27-4a28-af18-1f93272d4d3c/2894-0
00:14:10.989 --> 00:14:13.667
It's like a single narrow
network, but we have different

025a3067-1e27-4a28-af18-1f93272d4d3c/2894-1
00:14:13.667 --> 00:14:16.392
passes inside of the neural
network and we presented this

025a3067-1e27-4a28-af18-1f93272d4d3c/2894-2
00:14:16.392 --> 00:14:18.929
amount of passes in in, in the
beginning of training.

025a3067-1e27-4a28-af18-1f93272d4d3c/2902-0
00:14:19.019 --> 00:14:22.886
We can say, well, we're gonna
have 128 passes in each of the

025a3067-1e27-4a28-af18-1f93272d4d3c/2902-1
00:14:22.886 --> 00:14:23.329
layers.

025a3067-1e27-4a28-af18-1f93272d4d3c/2910-0
00:14:23.779 --> 00:14:25.509
Each is maximal layers.

025a3067-1e27-4a28-af18-1f93272d4d3c/2915-0
00:14:26.079 --> 00:14:28.569
All right, bear with me.

025a3067-1e27-4a28-af18-1f93272d4d3c/2920-0
00:14:28.819 --> 00:14:30.129
Gonna get easier.

025a3067-1e27-4a28-af18-1f93272d4d3c/2927-0
00:14:31.099 --> 00:14:32.209
So we have an input.

025a3067-1e27-4a28-af18-1f93272d4d3c/2946-0
00:14:33.319 --> 00:14:37.709
Uh, we would make a copy of this
input and batch it together.

025a3067-1e27-4a28-af18-1f93272d4d3c/3000-0
00:14:37.999 --> 00:14:41.913
So instead of, you know, copying
an error network or instead of

025a3067-1e27-4a28-af18-1f93272d4d3c/3000-1
00:14:41.913 --> 00:14:45.643
running it multiple times, we
create a copy of the input and

025a3067-1e27-4a28-af18-1f93272d4d3c/3000-2
00:14:45.643 --> 00:14:49.251
we run this input through an
error network that have these

025a3067-1e27-4a28-af18-1f93272d4d3c/3000-3
00:14:49.251 --> 00:14:52.369
different subnetworks you know,
passes through it.

025a3067-1e27-4a28-af18-1f93272d4d3c/3037-0
00:14:53.539 --> 00:14:56.975
Then each of these each of these
sub network would make a

025a3067-1e27-4a28-af18-1f93272d4d3c/3037-1
00:14:56.975 --> 00:15:00.766
prediction and those predictions
will be slightly different and

025a3067-1e27-4a28-af18-1f93272d4d3c/3037-2
00:15:00.766 --> 00:15:04.498
then we can aggregate them and
estimate the uncertainty of the

025a3067-1e27-4a28-af18-1f93272d4d3c/3037-3
00:15:04.498 --> 00:15:05.149
prediction.

025a3067-1e27-4a28-af18-1f93272d4d3c/3052-0
00:15:05.219 --> 00:15:08.889
And this is great because you
have a single narrow network.

025a3067-1e27-4a28-af18-1f93272d4d3c/3075-0
00:15:08.899 --> 00:15:13.294
You have a single input and it's
all sort of trained in the way

025a3067-1e27-4a28-af18-1f93272d4d3c/3075-1
00:15:13.294 --> 00:15:15.079
to give you a good answer.

025a3067-1e27-4a28-af18-1f93272d4d3c/3086-0
00:15:15.959 --> 00:15:18.609
When we train it, we don't have
a notion of probability.

025a3067-1e27-4a28-af18-1f93272d4d3c/3098-0
00:15:18.619 --> 00:15:20.728
We don't have a notion of
uncertainty or anything like

025a3067-1e27-4a28-af18-1f93272d4d3c/3098-1
00:15:20.728 --> 00:15:20.919
that.

025a3067-1e27-4a28-af18-1f93272d4d3c/3109-0
00:15:20.929 --> 00:15:23.469
We just train it as we would
usually do.

025a3067-1e27-4a28-af18-1f93272d4d3c/3153-0
00:15:23.579 --> 00:15:27.956
We just with samples predicting
values, but then as soon as we

025a3067-1e27-4a28-af18-1f93272d4d3c/3153-1
00:15:27.956 --> 00:15:32.262
have an inference part when we
predict something based on our

025a3067-1e27-4a28-af18-1f93272d4d3c/3153-2
00:15:32.262 --> 00:15:36.569
data we have uh, we're able to
predict the uncertainty of the

025a3067-1e27-4a28-af18-1f93272d4d3c/3153-3
00:15:36.569 --> 00:15:37.819
prediction itself.

025a3067-1e27-4a28-af18-1f93272d4d3c/3159-0
00:15:38.899 --> 00:15:39.279
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/3166-0
00:15:39.289 --> 00:15:41.929
Which is a powerful idea, umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/3224-0
00:15:42.499 --> 00:15:47.033
And because we use a regression
instead of classifying 3000

025a3067-1e27-4a28-af18-1f93272d4d3c/3224-1
00:15:47.033 --> 00:15:51.113
samples, it's easy for us
because we only predict you

025a3067-1e27-4a28-af18-1f93272d4d3c/3224-2
00:15:51.113 --> 00:15:55.798
know, 2 numbers, we don't need
to carry like 128 copies of an

025a3067-1e27-4a28-af18-1f93272d4d3c/3224-3
00:15:55.798 --> 00:15:58.669
array of like 3 by 3000 or 3 by
6000.

025a3067-1e27-4a28-af18-1f93272d4d3c/3230-0
00:15:58.749 --> 00:16:00.689
And it scales even further.

025a3067-1e27-4a28-af18-1f93272d4d3c/3262-0
00:16:00.699 --> 00:16:04.578
If you have a window of like
15,000 samples and then you have

025a3067-1e27-4a28-af18-1f93272d4d3c/3262-1
00:16:04.578 --> 00:16:08.519
to have a copy, you could see
how the the amount of parameters

025a3067-1e27-4a28-af18-1f93272d4d3c/3262-2
00:16:08.519 --> 00:16:09.519
growth for that.

025a3067-1e27-4a28-af18-1f93272d4d3c/3279-0
00:16:09.879 --> 00:16:14.640
So Speaking of results being
curious here, one example of a

025a3067-1e27-4a28-af18-1f93272d4d3c/3279-1
00:16:14.640 --> 00:16:15.989
prediction right.

025a3067-1e27-4a28-af18-1f93272d4d3c/3307-0
00:16:16.059 --> 00:16:20.686
So we have our waveform here and
we are predicting AP&S wave

025a3067-1e27-4a28-af18-1f93272d4d3c/3307-1
00:16:20.686 --> 00:16:22.109
and red and in blue.

025a3067-1e27-4a28-af18-1f93272d4d3c/3321-0
00:16:22.179 --> 00:16:24.729
And here is just a zoom in so we
can see better.

025a3067-1e27-4a28-af18-1f93272d4d3c/3354-0
00:16:26.009 --> 00:16:30.157
So the dashed line is where it
should be and where the

025a3067-1e27-4a28-af18-1f93272d4d3c/3354-1
00:16:30.157 --> 00:16:34.984
prediction should be and the the
solid opaque line is where the

025a3067-1e27-4a28-af18-1f93272d4d3c/3354-2
00:16:34.984 --> 00:16:36.039
prediction is.

025a3067-1e27-4a28-af18-1f93272d4d3c/3377-0
00:16:36.169 --> 00:16:40.635
And if you look closely enough,
then you're gonna see it's gonna

025a3067-1e27-4a28-af18-1f93272d4d3c/3377-1
00:16:40.635 --> 00:16:43.039
be more evident on the next
slide.

025a3067-1e27-4a28-af18-1f93272d4d3c/3427-0
00:16:43.209 --> 00:16:46.690
The distribution of probability
is not necessarily Gaussian,

025a3067-1e27-4a28-af18-1f93272d4d3c/3427-1
00:16:46.690 --> 00:16:50.284
like we can approximate it as
Gaussian, but it doesn't have to

025a3067-1e27-4a28-af18-1f93272d4d3c/3427-2
00:16:50.284 --> 00:16:53.764
be Gaussian like you just have
actual values for each of the

025a3067-1e27-4a28-af18-1f93272d4d3c/3427-3
00:16:53.764 --> 00:16:57.301
predictions and then you know
whatever distribution they form

025a3067-1e27-4a28-af18-1f93272d4d3c/3427-4
00:16:57.301 --> 00:16:57.529
you.

025a3067-1e27-4a28-af18-1f93272d4d3c/3430-0
00:16:57.539 --> 00:16:58.149
You got it.

025a3067-1e27-4a28-af18-1f93272d4d3c/3437-0
00:16:58.239 --> 00:16:59.689
So example given here.

025a3067-1e27-4a28-af18-1f93272d4d3c/3460-0
00:16:59.779 --> 00:17:02.990
Like it seems that there was
like a second potential

025a3067-1e27-4a28-af18-1f93272d4d3c/3460-1
00:17:02.990 --> 00:17:06.746
candidate where the phase might
have started and so on and so

025a3067-1e27-4a28-af18-1f93272d4d3c/3460-2
00:17:06.746 --> 00:17:07.109
forth.

025a3067-1e27-4a28-af18-1f93272d4d3c/3478-0
00:17:07.639 --> 00:17:12.540
Now the biggest problem with
that is that the data we have is

025a3067-1e27-4a28-af18-1f93272d4d3c/3478-1
00:17:12.540 --> 00:17:13.409
very noisy.

025a3067-1e27-4a28-af18-1f93272d4d3c/3507-0
00:17:13.799 --> 00:17:17.936
All of the picks you know, when
I was studying Moscow state, we

025a3067-1e27-4a28-af18-1f93272d4d3c/3507-1
00:17:17.936 --> 00:17:21.879
used to say to geologists 3
opinions and same with smallest.

025a3067-1e27-4a28-af18-1f93272d4d3c/3519-0
00:17:21.889 --> 00:17:25.589
You know, I bet like all of us
are experts, you know, trained.

025a3067-1e27-4a28-af18-1f93272d4d3c/3533-0
00:17:25.599 --> 00:17:27.349
But you know, give us the same
wave forms.

025a3067-1e27-4a28-af18-1f93272d4d3c/3546-0
00:17:27.359 --> 00:17:30.638
We're gonna pick different and
and that's the data that we're

025a3067-1e27-4a28-af18-1f93272d4d3c/3546-1
00:17:30.638 --> 00:17:31.219
working on.

025a3067-1e27-4a28-af18-1f93272d4d3c/3566-0
00:17:31.289 --> 00:17:35.853
So we have this inherit sort of
uncertainty and noise in the

025a3067-1e27-4a28-af18-1f93272d4d3c/3566-1
00:17:35.853 --> 00:17:37.199
data, which makes.

025a3067-1e27-4a28-af18-1f93272d4d3c/3602-0
00:17:37.769 --> 00:17:41.397
So with that in mind, it makes a
lot of sense to kind of estimate

025a3067-1e27-4a28-af18-1f93272d4d3c/3602-1
00:17:41.397 --> 00:17:44.311
the uncertainty of your
prediction and then use this

025a3067-1e27-4a28-af18-1f93272d4d3c/3602-2
00:17:44.311 --> 00:17:46.949
uncertainty for whatever you do
with this data.

025a3067-1e27-4a28-af18-1f93272d4d3c/3613-0
00:17:46.959 --> 00:17:49.869
Like you know, inversion and so
on and so forth.

025a3067-1e27-4a28-af18-1f93272d4d3c/3616-0
00:17:50.479 --> 00:17:50.899
Right.

025a3067-1e27-4a28-af18-1f93272d4d3c/3653-0
00:17:50.989 --> 00:17:55.820
So what we have so far is that
we have a neural network that is

025a3067-1e27-4a28-af18-1f93272d4d3c/3653-1
00:17:55.820 --> 00:18:00.575
able to pick PNS phase and is
able to estimate the uncertainty

025a3067-1e27-4a28-af18-1f93272d4d3c/3653-2
00:18:00.575 --> 00:18:02.009
of this prediction.

025a3067-1e27-4a28-af18-1f93272d4d3c/3677-0
00:18:04.129 --> 00:18:07.629
And if we were to compare it to
other methods, it's gonna be a

025a3067-1e27-4a28-af18-1f93272d4d3c/3677-1
00:18:07.629 --> 00:18:08.629
very boring slide.

025a3067-1e27-4a28-af18-1f93272d4d3c/3689-0
00:18:08.639 --> 00:18:12.176
But you'd see that face Hunter
beats phase net and all other

025a3067-1e27-4a28-af18-1f93272d4d3c/3689-1
00:18:12.176 --> 00:18:12.639
methods.

025a3067-1e27-4a28-af18-1f93272d4d3c/3693-0
00:18:13.269 --> 00:18:13.889
Umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/3703-0
00:18:14.339 --> 00:18:17.869
On different data sets I like
here is an example.

025a3067-1e27-4a28-af18-1f93272d4d3c/3713-0
00:18:17.939 --> 00:18:19.889
Test it on stead and here is an
example.

025a3067-1e27-4a28-af18-1f93272d4d3c/3752-0
00:18:19.899 --> 00:18:25.026
Tested on ETH C and if you look
at the the values and like P

025a3067-1e27-4a28-af18-1f93272d4d3c/3752-1
00:18:25.026 --> 00:18:30.321
error values and S error values,
you'd see that there are much

025a3067-1e27-4a28-af18-1f93272d4d3c/3752-2
00:18:30.321 --> 00:18:30.909
better.

025a3067-1e27-4a28-af18-1f93272d4d3c/3782-0
00:18:30.999 --> 00:18:35.919
But again, honestly, I don't
think that the difference of

025a3067-1e27-4a28-af18-1f93272d4d3c/3782-1
00:18:35.919 --> 00:18:40.159
00.4 and 0.1 like makes a real
difference for us.

025a3067-1e27-4a28-af18-1f93272d4d3c/3814-0
00:18:40.169 --> 00:18:43.802
You know, because it doesn't
translate to, like huge

025a3067-1e27-4a28-af18-1f93272d4d3c/3814-1
00:18:43.802 --> 00:18:47.914
distances errors or anything
like this, but it's, you know,

025a3067-1e27-4a28-af18-1f93272d4d3c/3814-2
00:18:47.914 --> 00:18:49.559
it's great to have, umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/3830-0
00:18:50.069 --> 00:18:55.744
Then another interesting thing
to observe is when you have your

025a3067-1e27-4a28-af18-1f93272d4d3c/3830-1
00:18:55.744 --> 00:18:58.049
neural network predicting.

025a3067-1e27-4a28-af18-1f93272d4d3c/3832-0
00:18:58.959 --> 00:18:59.439
Umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/3880-0
00:19:00.669 --> 00:19:05.421
And uncertainty, right, you as
you add more noise to this data,

025a3067-1e27-4a28-af18-1f93272d4d3c/3880-1
00:19:05.421 --> 00:19:10.320
the more sort of uncertainty you
have in this prediction, this is

025a3067-1e27-4a28-af18-1f93272d4d3c/3880-2
00:19:10.320 --> 00:19:14.700
like a simple test to see
whether your neural net is doing

025a3067-1e27-4a28-af18-1f93272d4d3c/3880-3
00:19:14.700 --> 00:19:16.259
something reasonable.

025a3067-1e27-4a28-af18-1f93272d4d3c/3888-0
00:19:16.349 --> 00:19:18.059
Again, we have a a waveform.

025a3067-1e27-4a28-af18-1f93272d4d3c/3902-0
00:19:18.069 --> 00:19:21.309
I'm just adding Gaussian noise
to it and we can see that the.

025a3067-1e27-4a28-af18-1f93272d4d3c/3918-0
00:19:23.319 --> 00:19:28.122
Uncertainty is increasing for
those predictions as the noise

025a3067-1e27-4a28-af18-1f93272d4d3c/3918-1
00:19:28.122 --> 00:19:29.539
levels are rising.

025a3067-1e27-4a28-af18-1f93272d4d3c/3954-0
00:19:30.179 --> 00:19:34.113
But what's actually interesting
to observe is the correlation

025a3067-1e27-4a28-af18-1f93272d4d3c/3954-1
00:19:34.113 --> 00:19:37.729
between predicted and certainty
and mean absolute error.

025a3067-1e27-4a28-af18-1f93272d4d3c/4071-0
00:19:38.119 --> 00:19:43.148
If you look at the slide, you'll
see that on the bottom here we

025a3067-1e27-4a28-af18-1f93272d4d3c/4071-1
00:19:43.148 --> 00:19:47.942
have the uncertainty and on the
on the left here we have the

025a3067-1e27-4a28-af18-1f93272d4d3c/4071-2
00:19:47.942 --> 00:19:52.892
mean absolute error and if and
blue and red are for PNS peaks,

025a3067-1e27-4a28-af18-1f93272d4d3c/4071-3
00:19:52.892 --> 00:19:57.764
right and we see that if the
error like if the uncertainty is

025a3067-1e27-4a28-af18-1f93272d4d3c/4071-4
00:19:57.764 --> 00:20:02.478
small and the error is also
small and it works if we have a

025a3067-1e27-4a28-af18-1f93272d4d3c/4071-5
00:20:02.478 --> 00:20:07.350
very high uncertainty, the error
is likely to be very high as

025a3067-1e27-4a28-af18-1f93272d4d3c/4071-6
00:20:07.350 --> 00:20:12.065
well, which means that in a
sense we're starting to be able

025a3067-1e27-4a28-af18-1f93272d4d3c/4071-7
00:20:12.065 --> 00:20:16.779
to have an error network that
kind of knows when it's wrong

025a3067-1e27-4a28-af18-1f93272d4d3c/4071-8
00:20:16.779 --> 00:20:19.529
like it's not necessarily like
it.

025a3067-1e27-4a28-af18-1f93272d4d3c/4112-0
00:20:19.539 --> 00:20:22.899
There are lots of lots of
evaluations there needed still

025a3067-1e27-4a28-af18-1f93272d4d3c/4112-1
00:20:22.899 --> 00:20:26.612
to to be like absolutely solid
and say like, yeah, we actually

025a3067-1e27-4a28-af18-1f93272d4d3c/4112-2
00:20:26.612 --> 00:20:29.559
are able to like cut off all the
bad predictions.

025a3067-1e27-4a28-af18-1f93272d4d3c/4119-0
00:20:29.629 --> 00:20:30.859
But it's a great start, I think.

025a3067-1e27-4a28-af18-1f93272d4d3c/4146-0
00:20:31.299 --> 00:20:35.706
And that we have this ability to
estimate the the error and you

025a3067-1e27-4a28-af18-1f93272d4d3c/4146-1
00:20:35.706 --> 00:20:39.079
know here is an example of how
you might use it.

025a3067-1e27-4a28-af18-1f93272d4d3c/4156-0
00:20:39.709 --> 00:20:41.639
You, let's say you have a
network.

025a3067-1e27-4a28-af18-1f93272d4d3c/4178-0
00:20:41.649 --> 00:20:45.264
You you're picking on this
network and then as the

025a3067-1e27-4a28-af18-1f93272d4d3c/4178-1
00:20:45.264 --> 00:20:49.019
confidence in the prediction
fades, so does the pig.

025a3067-1e27-4a28-af18-1f93272d4d3c/4201-0
00:20:49.769 --> 00:20:53.605
So we start with like a very
solid text here and as the

025a3067-1e27-4a28-af18-1f93272d4d3c/4201-1
00:20:53.605 --> 00:20:56.619
distance grows, you know the IT
fades away.

025a3067-1e27-4a28-af18-1f93272d4d3c/4303-0
00:20:56.629 --> 00:20:59.688
The the reason for this
particular example is because I

025a3067-1e27-4a28-af18-1f93272d4d3c/4303-1
00:20:59.688 --> 00:21:03.128
trained it on a regional data
and then I was giving it further

025a3067-1e27-4a28-af18-1f93272d4d3c/4303-2
00:21:03.128 --> 00:21:06.296
away distances just to see
because they expect it will be

025a3067-1e27-4a28-af18-1f93272d4d3c/4303-3
00:21:06.296 --> 00:21:09.737
that if the distance is far away
then you know it shouldn't be

025a3067-1e27-4a28-af18-1f93272d4d3c/4303-4
00:21:09.737 --> 00:21:13.341
certain because we didn't really
have it in the training data and

025a3067-1e27-4a28-af18-1f93272d4d3c/4303-5
00:21:13.341 --> 00:21:16.618
it works right and then you can
like when you're estimating

025a3067-1e27-4a28-af18-1f93272d4d3c/4303-6
00:21:16.618 --> 00:21:20.168
velocity or something like this
you can use those confidences as

025a3067-1e27-4a28-af18-1f93272d4d3c/4303-7
00:21:20.168 --> 00:21:23.663
weights to your whatever you do
regression version or something

025a3067-1e27-4a28-af18-1f93272d4d3c/4303-8
00:21:23.663 --> 00:21:24.209
like this.

025a3067-1e27-4a28-af18-1f93272d4d3c/4319-0
00:21:24.259 --> 00:21:27.399
And now you're you're able to
get like a more accurate

025a3067-1e27-4a28-af18-1f93272d4d3c/4319-1
00:21:27.399 --> 00:21:27.969
responses.

025a3067-1e27-4a28-af18-1f93272d4d3c/4338-0
00:21:29.519 --> 00:21:33.795
And as a bonus, with Face
Hunter, you get shift in

025a3067-1e27-4a28-af18-1f93272d4d3c/4338-1
00:21:33.795 --> 00:21:34.549
variance.

025a3067-1e27-4a28-af18-1f93272d4d3c/4389-0
00:21:34.619 --> 00:21:38.779
So if you try to do the same
experience with facenet, you're

025a3067-1e27-4a28-af18-1f93272d4d3c/4389-1
00:21:38.779 --> 00:21:43.007
gonna see that the prediction is
gonna be jumping quite a lot

025a3067-1e27-4a28-af18-1f93272d4d3c/4389-2
00:21:43.007 --> 00:21:47.303
because as you transition your
samples through your window and

025a3067-1e27-4a28-af18-1f93272d4d3c/4389-3
00:21:47.303 --> 00:21:51.599
phase net is is giving you like
slightly different predictions

025a3067-1e27-4a28-af18-1f93272d4d3c/4389-4
00:21:51.599 --> 00:21:52.349
every time.

025a3067-1e27-4a28-af18-1f93272d4d3c/4437-0
00:21:52.519 --> 00:21:55.098
The reason for that is not
because phase net is bad or

025a3067-1e27-4a28-af18-1f93272d4d3c/4437-1
00:21:55.098 --> 00:21:55.989
anything like this.

025a3067-1e27-4a28-af18-1f93272d4d3c/4495-0
00:21:56.119 --> 00:21:59.838
Is the the way how convolution
is done and it was just a

025a3067-1e27-4a28-af18-1f93272d4d3c/4495-1
00:21:59.838 --> 00:22:03.492
traditional way of how
convolutions were applied and at

025a3067-1e27-4a28-af18-1f93272d4d3c/4495-2
00:22:03.492 --> 00:22:07.341
the time and which introduced
aliasing in 2019 there was a

025a3067-1e27-4a28-af18-1f93272d4d3c/4495-3
00:22:07.341 --> 00:22:10.538
paper called making
convolutional networks shift

025a3067-1e27-4a28-af18-1f93272d4d3c/4495-4
00:22:10.538 --> 00:22:14.779
invariant again, and this paper
demonstrated how to just to have

025a3067-1e27-4a28-af18-1f93272d4d3c/4495-5
00:22:14.779 --> 00:22:18.824
a drop in replacement for the
neural network that would allow

025a3067-1e27-4a28-af18-1f93272d4d3c/4495-6
00:22:18.824 --> 00:22:22.869
you to sort of get rid of shift
variance in your predictions.

025a3067-1e27-4a28-af18-1f93272d4d3c/4534-0
00:22:22.879 --> 00:22:26.226
It was done for images
originally like the example they

025a3067-1e27-4a28-af18-1f93272d4d3c/4534-1
00:22:26.226 --> 00:22:29.991
have is an image of a doc and
then as you drag this image of a

025a3067-1e27-4a28-af18-1f93272d4d3c/4534-2
00:22:29.991 --> 00:22:32.739
dog, the classification starts
to vary a lot.

025a3067-1e27-4a28-af18-1f93272d4d3c/4538-0
00:22:32.749 --> 00:22:33.819
Like, sometimes it's a dog.

025a3067-1e27-4a28-af18-1f93272d4d3c/4557-0
00:22:33.959 --> 00:22:37.152
Then you move it like few pixels
left and then suddenly not the

025a3067-1e27-4a28-af18-1f93272d4d3c/4557-1
00:22:37.152 --> 00:22:38.149
dock anymore, right?

025a3067-1e27-4a28-af18-1f93272d4d3c/4565-0
00:22:38.159 --> 00:22:38.949
It doesn't make sense.

025a3067-1e27-4a28-af18-1f93272d4d3c/4625-0
00:22:40.419 --> 00:22:44.414
I in general, I think it's a
very useful sort of technique to

025a3067-1e27-4a28-af18-1f93272d4d3c/4625-1
00:22:44.414 --> 00:22:48.602
think about our data in terms of
images, because images are easy

025a3067-1e27-4a28-af18-1f93272d4d3c/4625-2
00:22:48.602 --> 00:22:52.211
to visualize, it's easy to
visualize a doc right on the

025a3067-1e27-4a28-af18-1f93272d4d3c/4625-3
00:22:52.211 --> 00:22:56.012
photo, and it's much more
difficult to visualize it P face

025a3067-1e27-4a28-af18-1f93272d4d3c/4625-4
00:22:56.012 --> 00:22:58.009
in some high dimensional space.

025a3067-1e27-4a28-af18-1f93272d4d3c/4697-0
00:22:58.139 --> 00:23:01.954
But the underlying principles in
the string as the same as we can

025a3067-1e27-4a28-af18-1f93272d4d3c/4697-1
00:23:01.954 --> 00:23:05.365
see now with large language
models and Transformers and so

025a3067-1e27-4a28-af18-1f93272d4d3c/4697-2
00:23:05.365 --> 00:23:08.833
on and so forth, we are just
using the same neural networks

025a3067-1e27-4a28-af18-1f93272d4d3c/4697-3
00:23:08.833 --> 00:23:12.532
were different types of data and
it just works like we're using

025a3067-1e27-4a28-af18-1f93272d4d3c/4697-4
00:23:12.532 --> 00:23:15.711
the same techniques for
different types of data and it

025a3067-1e27-4a28-af18-1f93272d4d3c/4697-5
00:23:15.711 --> 00:23:16.809
just works alright.

025a3067-1e27-4a28-af18-1f93272d4d3c/4716-0
00:23:17.729 --> 00:23:23.146
But now the part that I promised
uh, going beyond face speaking,

025a3067-1e27-4a28-af18-1f93272d4d3c/4716-1
00:23:23.146 --> 00:23:24.479
probably by now.

025a3067-1e27-4a28-af18-1f93272d4d3c/4750-0
00:23:24.489 --> 00:23:28.448
You kind of have an idea of how
to do that and imagine you have

025a3067-1e27-4a28-af18-1f93272d4d3c/4750-1
00:23:28.448 --> 00:23:32.282
a training set where you try to
predict some other parameters

025a3067-1e27-4a28-af18-1f93272d4d3c/4750-2
00:23:32.282 --> 00:23:33.519
rather than P phase.

025a3067-1e27-4a28-af18-1f93272d4d3c/4765-0
00:23:33.789 --> 00:23:37.823
You can try to predict you know
10 regional phases in the same

025a3067-1e27-4a28-af18-1f93272d4d3c/4765-1
00:23:37.823 --> 00:23:38.399
paradigm.

025a3067-1e27-4a28-af18-1f93272d4d3c/4797-0
00:23:38.409 --> 00:23:41.982
You have like a linear layer in
the end and you have 10 outputs

025a3067-1e27-4a28-af18-1f93272d4d3c/4797-1
00:23:41.982 --> 00:23:42.819
instead of two.

025a3067-1e27-4a28-af18-1f93272d4d3c/4830-0
00:23:43.029 --> 00:23:46.707
And if you have a good enough
training data, you're gonna have

025a3067-1e27-4a28-af18-1f93272d4d3c/4830-1
00:23:46.707 --> 00:23:50.267
regional phases predicted and
you can enlarge the window and

025a3067-1e27-4a28-af18-1f93272d4d3c/4830-2
00:23:50.267 --> 00:23:53.769
you wouldn't need to change
anything in the network because

025a3067-1e27-4a28-af18-1f93272d4d3c/4830-3
00:23:53.769 --> 00:23:56.629
we're always having from one to
zero in percent.

025a3067-1e27-4a28-af18-1f93272d4d3c/4840-0
00:23:56.769 --> 00:23:58.139
So it's very easy to set up.

025a3067-1e27-4a28-af18-1f93272d4d3c/4858-0
00:23:58.149 --> 00:24:01.553
It's very easy to start using it
straight away and train it on on

025a3067-1e27-4a28-af18-1f93272d4d3c/4858-1
00:24:01.553 --> 00:24:02.429
a different data.

025a3067-1e27-4a28-af18-1f93272d4d3c/4862-0
00:24:03.509 --> 00:24:03.879
Umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/4887-0
00:24:05.459 --> 00:24:08.957
So if you are trying to go
beyond face taking, we can try

025a3067-1e27-4a28-af18-1f93272d4d3c/4887-1
00:24:08.957 --> 00:24:11.609
to predict some properties of
the waveform.

025a3067-1e27-4a28-af18-1f93272d4d3c/4900-0
00:24:11.899 --> 00:24:16.444
Let's say we have earthquakes
recorded and we're trying to

025a3067-1e27-4a28-af18-1f93272d4d3c/4900-1
00:24:16.444 --> 00:24:17.599
predict depths.

025a3067-1e27-4a28-af18-1f93272d4d3c/4923-0
00:24:18.709 --> 00:24:21.929
Then just to make it more
challenging for the neural

025a3067-1e27-4a28-af18-1f93272d4d3c/4923-1
00:24:21.929 --> 00:24:24.359
network, instead of using 3000
samples.

025a3067-1e27-4a28-af18-1f93272d4d3c/4930-0
00:24:24.369 --> 00:24:26.179
Yeah, let's just use 100
samples.

025a3067-1e27-4a28-af18-1f93272d4d3c/4999-0
00:24:26.369 --> 00:24:29.747
Let's just say we're trying to
do it in a real time and we

025a3067-1e27-4a28-af18-1f93272d4d3c/4999-1
00:24:29.747 --> 00:24:33.296
wanna have as little window as
possible so we don't have like

025a3067-1e27-4a28-af18-1f93272d4d3c/4999-2
00:24:33.296 --> 00:24:36.674
that much latency between our
predictions because the good

025a3067-1e27-4a28-af18-1f93272d4d3c/4999-3
00:24:36.674 --> 00:24:40.395
thing about neural networks that
the prediction itself is almost

025a3067-1e27-4a28-af18-1f93272d4d3c/4999-4
00:24:40.395 --> 00:24:43.887
instant like you have like 0.2
milliseconds for a prediction

025a3067-1e27-4a28-af18-1f93272d4d3c/4999-5
00:24:43.887 --> 00:24:45.489
and it's not even optimized.

025a3067-1e27-4a28-af18-1f93272d4d3c/5011-0
00:24:45.499 --> 00:24:47.919
Like you can optimize it even
further and further and further.

025a3067-1e27-4a28-af18-1f93272d4d3c/5027-0
00:24:48.029 --> 00:24:52.339
So then the only latency you
have is the window you're

025a3067-1e27-4a28-af18-1f93272d4d3c/5027-1
00:24:52.339 --> 00:24:53.749
dragging with you.

025a3067-1e27-4a28-af18-1f93272d4d3c/5054-0
00:24:53.799 --> 00:24:58.024
So I just experimented with that
and you can take one second of

025a3067-1e27-4a28-af18-1f93272d4d3c/5054-1
00:24:58.024 --> 00:25:02.117
the data where you have IP face
somewhere like roughly in the

025a3067-1e27-4a28-af18-1f93272d4d3c/5054-2
00:25:02.117 --> 00:25:02.579
middle.

025a3067-1e27-4a28-af18-1f93272d4d3c/5069-0
00:25:03.549 --> 00:25:06.099
And so you had to detect her and
the detector to face.

025a3067-1e27-4a28-af18-1f93272d4d3c/5090-0
00:25:06.109 --> 00:25:09.403
So now you're taking this this
frame and you can start to

025a3067-1e27-4a28-af18-1f93272d4d3c/5090-1
00:25:09.403 --> 00:25:11.219
predict depths of an earthquake.

025a3067-1e27-4a28-af18-1f93272d4d3c/5118-0
00:25:11.859 --> 00:25:16.093
And it works surprisingly
weirdly well, like it doesn't

025a3067-1e27-4a28-af18-1f93272d4d3c/5118-1
00:25:16.093 --> 00:25:20.176
work like ideally, but neither
was that data set like

025a3067-1e27-4a28-af18-1f93272d4d3c/5118-2
00:25:20.176 --> 00:25:21.839
specifically for that.

025a3067-1e27-4a28-af18-1f93272d4d3c/5139-0
00:25:21.949 --> 00:25:25.553
I just took stat, you know which
which is not really verified for

025a3067-1e27-4a28-af18-1f93272d4d3c/5139-1
00:25:25.553 --> 00:25:27.299
depths and you know all of that.

025a3067-1e27-4a28-af18-1f93272d4d3c/5166-0
00:25:27.309 --> 00:25:31.479
It was verified for her own set
phases because Mustafa actually

025a3067-1e27-4a28-af18-1f93272d4d3c/5166-1
00:25:31.479 --> 00:25:35.649
literally went manually through
the whole data set and checked.

025a3067-1e27-4a28-af18-1f93272d4d3c/5190-0
00:25:35.659 --> 00:25:38.241
You know whether they onset
makes makes sense, but they

025a3067-1e27-4a28-af18-1f93272d4d3c/5190-1
00:25:38.241 --> 00:25:40.729
didn't really do that with tabs
and other parameters.

025a3067-1e27-4a28-af18-1f93272d4d3c/5256-0
00:25:40.809 --> 00:25:43.983
The other thing you can try to
predict is the distance to the

025a3067-1e27-4a28-af18-1f93272d4d3c/5256-1
00:25:43.983 --> 00:25:46.748
earthquake and this is this
opens up a an interesting

025a3067-1e27-4a28-af18-1f93272d4d3c/5256-2
00:25:46.748 --> 00:25:49.666
direction in earthquake early
warning because if you can

025a3067-1e27-4a28-af18-1f93272d4d3c/5256-3
00:25:49.666 --> 00:25:53.044
forecast the distance like super
instantly and I'd like to remind

025a3067-1e27-4a28-af18-1f93272d4d3c/5256-4
00:25:53.044 --> 00:25:56.014
you that this is a single
station approach like you don't

025a3067-1e27-4a28-af18-1f93272d4d3c/5256-5
00:25:56.014 --> 00:25:57.549
even need an array of station.

025a3067-1e27-4a28-af18-1f93272d4d3c/5363-0
00:25:57.559 --> 00:26:00.702
It's like a single station, so
if you can forecast the

025a3067-1e27-4a28-af18-1f93272d4d3c/5363-1
00:26:00.702 --> 00:26:03.958
distance, everything like you
know, with some reasonable

025a3067-1e27-4a28-af18-1f93272d4d3c/5363-2
00:26:03.958 --> 00:26:07.443
accuracy, say if you're just
able to classify to near field,

025a3067-1e27-4a28-af18-1f93272d4d3c/5363-3
00:26:07.443 --> 00:26:11.157
you know far field near like far
away or something like this, it

025a3067-1e27-4a28-af18-1f93272d4d3c/5363-4
00:26:11.157 --> 00:26:14.528
already you know it's it's
already interesting because you

025a3067-1e27-4a28-af18-1f93272d4d3c/5363-5
00:26:14.528 --> 00:26:17.956
can use it for for so many
different things and and you can

025a3067-1e27-4a28-af18-1f93272d4d3c/5363-6
00:26:17.956 --> 00:26:21.326
try to work as magnitude but
magnitude is expected to work

025a3067-1e27-4a28-af18-1f93272d4d3c/5363-7
00:26:21.326 --> 00:26:24.811
because you know if you have
like enough samples you can see

025a3067-1e27-4a28-af18-1f93272d4d3c/5363-8
00:26:24.811 --> 00:26:27.439
the the the rapture in the sense
in the data.

025a3067-1e27-4a28-af18-1f93272d4d3c/5395-0
00:26:27.779 --> 00:26:31.409
But again, this is fun because
it's a single station method and

025a3067-1e27-4a28-af18-1f93272d4d3c/5395-1
00:26:31.409 --> 00:26:34.471
it gives you a reasonable
accuracy even without being

025a3067-1e27-4a28-af18-1f93272d4d3c/5395-2
00:26:34.471 --> 00:26:36.909
trained specifically on any of
the things.

025a3067-1e27-4a28-af18-1f93272d4d3c/5422-0
00:26:36.919 --> 00:26:41.519
So I bet you know you guys can
figure it out with nice data

025a3067-1e27-4a28-af18-1f93272d4d3c/5422-1
00:26:41.519 --> 00:26:44.969
sets and you know enough of time
to do that.

025a3067-1e27-4a28-af18-1f93272d4d3c/5435-0
00:26:45.039 --> 00:26:50.379
And here is a very solid example
on like a statistical example.

025a3067-1e27-4a28-af18-1f93272d4d3c/5452-0
00:26:50.469 --> 00:26:53.991
So what we have here is one
second, two second and three

025a3067-1e27-4a28-af18-1f93272d4d3c/5452-1
00:26:53.991 --> 00:26:54.979
seconds of data.

025a3067-1e27-4a28-af18-1f93272d4d3c/5474-0
00:26:55.489 --> 00:26:59.395
Here we have a a mean absolute
error like a distribution and

025a3067-1e27-4a28-af18-1f93272d4d3c/5474-1
00:26:59.395 --> 00:27:01.059
we're predicting distance.

025a3067-1e27-4a28-af18-1f93272d4d3c/5485-0
00:27:01.109 --> 00:27:03.219
So all of those plus are about
distance.

025a3067-1e27-4a28-af18-1f93272d4d3c/5544-0
00:27:03.269 --> 00:27:07.051
So as we can see, when we're
trying to predict on the test

025a3067-1e27-4a28-af18-1f93272d4d3c/5544-1
00:27:07.051 --> 00:27:11.153
set distance, we would have most
of the samples being predicted

025a3067-1e27-4a28-af18-1f93272d4d3c/5544-2
00:27:11.153 --> 00:27:15.254
correctly with the distances and
then you know it's sprats like

025a3067-1e27-4a28-af18-1f93272d4d3c/5544-3
00:27:15.254 --> 00:27:18.715
we have tails, but mostly like
it contained within 20

025a3067-1e27-4a28-af18-1f93272d4d3c/5544-4
00:27:18.715 --> 00:27:21.599
kilometers, 3 Michael, which is
interesting.

025a3067-1e27-4a28-af18-1f93272d4d3c/5569-0
00:27:21.689 --> 00:27:25.755
I think the maximum distance in
this data set was something like

025a3067-1e27-4a28-af18-1f93272d4d3c/5569-1
00:27:25.755 --> 00:27:28.319
200 kilometers, so you know a
10% error.

025a3067-1e27-4a28-af18-1f93272d4d3c/5571-0
00:27:28.329 --> 00:27:30.549
But it's interesting.

025a3067-1e27-4a28-af18-1f93272d4d3c/5587-0
00:27:30.599 --> 00:27:36.809
Uh, then S we give more data as
an input, we start to have.

025a3067-1e27-4a28-af18-1f93272d4d3c/5592-0
00:27:37.829 --> 00:27:38.249
You bet.

025a3067-1e27-4a28-af18-1f93272d4d3c/5626-0
00:27:38.259 --> 00:27:43.499
Are you know a better error like
here in the one second we have a

025a3067-1e27-4a28-af18-1f93272d4d3c/5626-1
00:27:43.499 --> 00:27:48.263
certain .8 kilometers absolute
error in 7.1 median absolute

025a3067-1e27-4a28-af18-1f93272d4d3c/5626-2
00:27:48.263 --> 00:27:48.739
error.

025a3067-1e27-4a28-af18-1f93272d4d3c/5655-0
00:27:48.969 --> 00:27:54.063
Then in case of two we have
11.75 point 3 and then you know

025a3067-1e27-4a28-af18-1f93272d4d3c/5655-1
00:27:54.063 --> 00:27:56.949
for three seconds 11.24 point
78.

025a3067-1e27-4a28-af18-1f93272d4d3c/5705-0
00:27:57.009 --> 00:28:00.556
So it improves and but the most
like the thing that was actually

025a3067-1e27-4a28-af18-1f93272d4d3c/5705-1
00:28:00.556 --> 00:28:03.939
surprised to see when I did it
for the first time because you

025a3067-1e27-4a28-af18-1f93272d4d3c/5705-2
00:28:03.939 --> 00:28:06.940
know some other people were
asking me to do that and I

025a3067-1e27-4a28-af18-1f93272d4d3c/5705-3
00:28:06.940 --> 00:28:10.159
didn't want to because I I
wasn't sure if it's gonna work.

025a3067-1e27-4a28-af18-1f93272d4d3c/5711-0
00:28:10.699 --> 00:28:12.329
I'm but here it is.

025a3067-1e27-4a28-af18-1f93272d4d3c/5759-0
00:28:12.339 --> 00:28:15.482
So we have a true distance
plotted versus predicted

025a3067-1e27-4a28-af18-1f93272d4d3c/5759-1
00:28:15.482 --> 00:28:19.289
distance and in color will have
a confidence in prediction and

025a3067-1e27-4a28-af18-1f93272d4d3c/5759-2
00:28:19.289 --> 00:28:22.916
the yellow it gets the more
confident the prediction is and

025a3067-1e27-4a28-af18-1f93272d4d3c/5759-3
00:28:22.916 --> 00:28:26.239
you can start to see that for
you know for one second.

025a3067-1e27-4a28-af18-1f93272d4d3c/5799-0
00:28:26.249 --> 00:28:30.134
It's like less pronounced, but
then for two seconds you see

025a3067-1e27-4a28-af18-1f93272d4d3c/5799-1
00:28:30.134 --> 00:28:33.566
this like yellow confident line
where the one to one

025a3067-1e27-4a28-af18-1f93272d4d3c/5799-2
00:28:33.566 --> 00:28:37.839
correspondence between predicted
distance and and and the the the

025a3067-1e27-4a28-af18-1f93272d4d3c/5799-3
00:28:37.839 --> 00:28:40.169
true distance saying for the
three.

025a3067-1e27-4a28-af18-1f93272d4d3c/5928-0
00:28:40.259 --> 00:28:44.204
So like in a sense, we're
starting like statistically,

025a3067-1e27-4a28-af18-1f93272d4d3c/5928-1
00:28:44.204 --> 00:28:48.652
we're starting to see that our
model is able to estimate when

025a3067-1e27-4a28-af18-1f93272d4d3c/5928-2
00:28:48.652 --> 00:28:52.740
something is wrong, like when it
doesn't really know the

025a3067-1e27-4a28-af18-1f93272d4d3c/5928-3
00:28:52.740 --> 00:28:57.259
prediction the way it works like
they did, what, like the the,

025a3067-1e27-4a28-af18-1f93272d4d3c/5928-4
00:28:57.259 --> 00:29:01.849
the reason why my work is when
you have your data like when you

025a3067-1e27-4a28-af18-1f93272d4d3c/5928-5
00:29:01.849 --> 00:29:06.440
have in there on network and you
train it on specific data, you

025a3067-1e27-4a28-af18-1f93272d4d3c/5928-6
00:29:06.440 --> 00:29:10.744
essentially are are building a
out of distribution detector

025a3067-1e27-4a28-af18-1f93272d4d3c/5928-7
00:29:10.744 --> 00:29:15.262
like if you had only the data
with earthquakes and never shown

025a3067-1e27-4a28-af18-1f93272d4d3c/5928-8
00:29:15.262 --> 00:29:19.136
any example of an explosion in
seismic data, then the

025a3067-1e27-4a28-af18-1f93272d4d3c/5928-9
00:29:19.136 --> 00:29:23.655
confidence for earthquakes is
going to be much higher than the

025a3067-1e27-4a28-af18-1f93272d4d3c/5928-10
00:29:23.655 --> 00:29:25.519
confidence for explosions.

025a3067-1e27-4a28-af18-1f93272d4d3c/5940-0
00:29:25.529 --> 00:29:28.089
Because just because you never
really demonstrated this data.

025a3067-1e27-4a28-af18-1f93272d4d3c/5965-0
00:29:28.099 --> 00:29:31.415
So the neural network is like
less certain than that and I

025a3067-1e27-4a28-af18-1f93272d4d3c/5965-1
00:29:31.415 --> 00:29:34.169
guess it translates to other
parameters as well.

025a3067-1e27-4a28-af18-1f93272d4d3c/5983-0
00:29:34.229 --> 00:29:38.205
Look, you know, as I mentioned,
there's a lot of research to be

025a3067-1e27-4a28-af18-1f93272d4d3c/5983-1
00:29:38.205 --> 00:29:39.199
done about that.

025a3067-1e27-4a28-af18-1f93272d4d3c/6021-0
00:29:40.529 --> 00:29:44.049
So as a conclusion, one can
safely replace counterintuitive

025a3067-1e27-4a28-af18-1f93272d4d3c/6021-1
00:29:44.049 --> 00:29:47.452
classification scheme with
regression, which is a concern

025a3067-1e27-4a28-af18-1f93272d4d3c/6021-2
00:29:47.452 --> 00:29:51.089
for you know, for people who are
building the neural network.

025a3067-1e27-4a28-af18-1f93272d4d3c/6033-0
00:29:51.099 --> 00:29:53.823
Classic concern of for people
who are using the neural

025a3067-1e27-4a28-af18-1f93272d4d3c/6033-1
00:29:53.823 --> 00:29:54.219
network.

025a3067-1e27-4a28-af18-1f93272d4d3c/6044-0
00:29:54.539 --> 00:29:57.769
This allows you to predict as
many parameters as needed.

025a3067-1e27-4a28-af18-1f93272d4d3c/6085-0
00:29:57.889 --> 00:30:01.912
So far I tested it on prediction
of six things at the same time

025a3067-1e27-4a28-af18-1f93272d4d3c/6085-1
00:30:01.912 --> 00:30:05.871
and I couldn't really see that
it deteriorates with the amount

025a3067-1e27-4a28-af18-1f93272d4d3c/6085-2
00:30:05.871 --> 00:30:09.139
of parameters, but again, it's a
very recent thing.

025a3067-1e27-4a28-af18-1f93272d4d3c/6173-0
00:30:09.689 --> 00:30:13.618
Hopefully more research would be
done about that, and in general

025a3067-1e27-4a28-af18-1f93272d4d3c/6173-1
00:30:13.618 --> 00:30:17.425
you know the regression approach
is very underexplored in deep

025a3067-1e27-4a28-af18-1f93272d4d3c/6173-2
00:30:17.425 --> 00:30:20.508
learning, not only in
seismology, but like in deep

025a3067-1e27-4a28-af18-1f93272d4d3c/6173-3
00:30:20.508 --> 00:30:24.134
learning in general, because
people usually don't deal with

025a3067-1e27-4a28-af18-1f93272d4d3c/6173-4
00:30:24.134 --> 00:30:27.639
seismic data or people don't
really deal with sensor data

025a3067-1e27-4a28-af18-1f93272d4d3c/6173-5
00:30:27.639 --> 00:30:31.387
that much and people are, you
know, are adept at using images

025a3067-1e27-4a28-af18-1f93272d4d3c/6173-6
00:30:31.387 --> 00:30:35.073
and all of that in your like
large language model is now are

025a3067-1e27-4a28-af18-1f93272d4d3c/6173-7
00:30:35.073 --> 00:30:37.309
thing and started started
yesterday.

025a3067-1e27-4a28-af18-1f93272d4d3c/6218-0
00:30:37.319 --> 00:30:40.896
The agents are now a big thing,
and we're gonna hear more and

025a3067-1e27-4a28-af18-1f93272d4d3c/6218-1
00:30:40.896 --> 00:30:44.184
more about that, but they're
regression approach is very

025a3067-1e27-4a28-af18-1f93272d4d3c/6218-2
00:30:44.184 --> 00:30:47.473
interesting because it allows
you to do many things that

025a3067-1e27-4a28-af18-1f93272d4d3c/6218-3
00:30:47.473 --> 00:30:49.549
weren't really possible without
it.

025a3067-1e27-4a28-af18-1f93272d4d3c/6243-0
00:30:50.409 --> 00:30:53.616
The and the other thing is that
you can use them uncertainty

025a3067-1e27-4a28-af18-1f93272d4d3c/6243-1
00:30:53.616 --> 00:30:56.559
measure to point, to quantify
confidence of your neural

025a3067-1e27-4a28-af18-1f93272d4d3c/6243-2
00:30:56.559 --> 00:30:56.979
network.

025a3067-1e27-4a28-af18-1f93272d4d3c/6272-0
00:30:57.149 --> 00:31:00.636
And this particular example in
phase speaking and then you can

025a3067-1e27-4a28-af18-1f93272d4d3c/6272-1
00:31:00.636 --> 00:31:03.569
use this uncertainty to as an
input to other things.

025a3067-1e27-4a28-af18-1f93272d4d3c/6292-0
00:31:04.539 --> 00:31:08.004
And you know, uncertainty, he's
also another unexplored area of

025a3067-1e27-4a28-af18-1f93272d4d3c/6292-1
00:31:08.004 --> 00:31:09.249
deep learning research.

025a3067-1e27-4a28-af18-1f93272d4d3c/6298-0
00:31:09.359 --> 00:31:11.489
But it's changing rapidly.

025a3067-1e27-4a28-af18-1f93272d4d3c/6350-0
00:31:12.179 --> 00:31:15.666
Like right now in Cambridge,
they're great courses on

025a3067-1e27-4a28-af18-1f93272d4d3c/6350-1
00:31:15.666 --> 00:31:19.734
predictive uncertainty for deep
learning and people are moving

025a3067-1e27-4a28-af18-1f93272d4d3c/6350-2
00:31:19.734 --> 00:31:23.609
beyond Bayesian research and
beyond Gaussian processes into

025a3067-1e27-4a28-af18-1f93272d4d3c/6350-3
00:31:23.609 --> 00:31:27.418
like this sort of like a weird
sampling approaches in deep

025a3067-1e27-4a28-af18-1f93272d4d3c/6350-4
00:31:27.418 --> 00:31:27.999
learning.

025a3067-1e27-4a28-af18-1f93272d4d3c/6382-0
00:31:28.009 --> 00:31:30.780
And I guess it's in the end of
the day, it's probably gonna be

025a3067-1e27-4a28-af18-1f93272d4d3c/6382-1
00:31:30.780 --> 00:31:33.594
a fusion of all of these things
together, like some smart PHD's

025a3067-1e27-4a28-af18-1f93272d4d3c/6382-2
00:31:33.594 --> 00:31:34.649
getting publish a paper.

025a3067-1e27-4a28-af18-1f93272d4d3c/6395-0
00:31:34.659 --> 00:31:37.324
And then everyone's like a year
later, everyone would be like,

025a3067-1e27-4a28-af18-1f93272d4d3c/6395-1
00:31:37.324 --> 00:31:38.339
oh, yeah, this is great.

025a3067-1e27-4a28-af18-1f93272d4d3c/6402-0
00:31:38.349 --> 00:31:39.609
Like, let's do that.

025a3067-1e27-4a28-af18-1f93272d4d3c/6406-0
00:31:40.769 --> 00:31:41.859
Umm yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/6437-0
00:31:41.869 --> 00:31:45.169
And it's also possible to update
their learning or in in a way

025a3067-1e27-4a28-af18-1f93272d4d3c/6437-1
00:31:45.169 --> 00:31:47.840
that doesn't introduce a
variance into your neural

025a3067-1e27-4a28-af18-1f93272d4d3c/6437-2
00:31:47.840 --> 00:31:48.259
network.

025a3067-1e27-4a28-af18-1f93272d4d3c/6444-0
00:31:48.649 --> 00:31:51.479
And now let me show you
something actually fun.

025a3067-1e27-4a28-af18-1f93272d4d3c/6454-0
00:31:53.069 --> 00:31:55.619
Umm how I'm doing in time?

025a3067-1e27-4a28-af18-1f93272d4d3c/6463-0
00:31:55.629 --> 00:31:59.359
How do I have what, 10 minutes
right 30.

025a3067-1e27-4a28-af18-1f93272d4d3c/6490-0
00:32:04.499 --> 00:32:09.602
So here's my GitHub right and
here's face Hunter and I just

025a3067-1e27-4a28-af18-1f93272d4d3c/6490-1
00:32:09.602 --> 00:32:13.939
want to demonstrate to you the
the wrapper itself.

025a3067-1e27-4a28-af18-1f93272d4d3c/6496-0
00:32:13.989 --> 00:32:15.299
So we have.

025a3067-1e27-4a28-af18-1f93272d4d3c/6506-0
00:32:15.499 --> 00:32:18.519
Here is a quick starter or quick
start guide.

025a3067-1e27-4a28-af18-1f93272d4d3c/6515-0
00:32:19.149 --> 00:32:21.719
You can, you know, you can
install the face hunter.

025a3067-1e27-4a28-af18-1f93272d4d3c/6533-0
00:32:21.729 --> 00:32:25.417
There's like a a whole
instruction on on how to install

025a3067-1e27-4a28-af18-1f93272d4d3c/6533-1
00:32:25.417 --> 00:32:26.339
a face hunter.

025a3067-1e27-4a28-af18-1f93272d4d3c/6548-0
00:32:26.609 --> 00:32:28.725
There is an instruction on how
to use the face hunter and

025a3067-1e27-4a28-af18-1f93272d4d3c/6548-1
00:32:28.725 --> 00:32:29.089
that's it.

025a3067-1e27-4a28-af18-1f93272d4d3c/6579-0
00:32:29.099 --> 00:32:32.734
Like this is all you need and
you just call the model and then

025a3067-1e27-4a28-af18-1f93272d4d3c/6579-1
00:32:32.734 --> 00:32:36.196
you call the prediction and you
can process continuous wave

025a3067-1e27-4a28-af18-1f93272d4d3c/6579-2
00:32:36.196 --> 00:32:37.349
forms straight away.

025a3067-1e27-4a28-af18-1f93272d4d3c/6595-0
00:32:37.359 --> 00:32:41.013
You can just, you know, upload
whatever waveforms you have, but

025a3067-1e27-4a28-af18-1f93272d4d3c/6595-1
00:32:41.013 --> 00:32:41.869
that's not fun.

025a3067-1e27-4a28-af18-1f93272d4d3c/6611-0
00:32:42.559 --> 00:32:47.119
This is fun, but no umm, why
not?

025a3067-1e27-4a28-af18-1f93272d4d3c/6615-0
00:32:48.949 --> 00:32:49.529
It was working.

025a3067-1e27-4a28-af18-1f93272d4d3c/6618-0
00:32:49.919 --> 00:32:51.249
Hmm, what?

025a3067-1e27-4a28-af18-1f93272d4d3c/6622-0
00:32:51.529 --> 00:32:52.779
Not fun for sure.

025a3067-1e27-4a28-af18-1f93272d4d3c/6626-0
00:32:57.859 --> 00:32:58.229
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/6640-0
00:32:58.239 --> 00:33:00.279
Give me the I guess it just
doesn't like.

025a3067-1e27-4a28-af18-1f93272d4d3c/6646-0
00:33:01.909 --> 00:33:03.809
Let me try use for me.

025a3067-1e27-4a28-af18-1f93272d4d3c/6681-0
00:33:07.559 --> 00:33:11.829
I didn't set it up because I
wasn't sure if we're gonna have

025a3067-1e27-4a28-af18-1f93272d4d3c/6681-1
00:33:11.829 --> 00:33:15.819
enough time, but this is like a
glimpse in the future of

025a3067-1e27-4a28-af18-1f93272d4d3c/6681-2
00:33:15.819 --> 00:33:17.079
probably Internet.

025a3067-1e27-4a28-af18-1f93272d4d3c/6687-0
00:33:24.719 --> 00:33:25.469
OK, here we are.

025a3067-1e27-4a28-af18-1f93272d4d3c/6691-0
00:33:42.939 --> 00:33:43.349
No.

025a3067-1e27-4a28-af18-1f93272d4d3c/6693-0
00:33:43.359 --> 00:33:43.669
OK.

025a3067-1e27-4a28-af18-1f93272d4d3c/6700-0
00:33:43.679 --> 00:33:44.669
Well, then it's gone.

025a3067-1e27-4a28-af18-1f93272d4d3c/6704-0
00:33:44.679 --> 00:33:46.829
Like someone deleted it. Umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/6712-0
00:33:51.979 --> 00:33:54.839
OK, well I tried. Umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/6717-0
00:33:58.669 --> 00:33:58.819
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/6724-0
00:33:58.829 --> 00:33:59.829
Yeah, I can show you how to do.

025a3067-1e27-4a28-af18-1f93272d4d3c/6726-0
00:33:59.839 --> 00:34:00.299
Uh.

025a3067-1e27-4a28-af18-1f93272d4d3c/6735-0
00:34:00.309 --> 00:34:02.299
How to how to do a new one?

025a3067-1e27-4a28-af18-1f93272d4d3c/6737-0
00:34:04.199 --> 00:34:04.479
Umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/6742-0
00:34:10.709 --> 00:34:11.149
OK.

025a3067-1e27-4a28-af18-1f93272d4d3c/6753-0
00:34:11.589 --> 00:34:15.009
I just the different account so
we can quickly create.

025a3067-1e27-4a28-af18-1f93272d4d3c/6756-0
00:34:16.669 --> 00:34:17.979
Nah, sorry.

025a3067-1e27-4a28-af18-1f93272d4d3c/6766-0
00:34:18.059 --> 00:34:19.339
OK, not gonna work.

025a3067-1e27-4a28-af18-1f93272d4d3c/6772-0
00:34:19.849 --> 00:34:20.699
Forget that then.

025a3067-1e27-4a28-af18-1f93272d4d3c/6779-0
00:34:20.749 --> 00:34:21.079
Alright.

025a3067-1e27-4a28-af18-1f93272d4d3c/6818-0
00:34:21.269 --> 00:34:26.752
Umm, I can show it to my my
laptop but it's give me now I

025a3067-1e27-4a28-af18-1f93272d4d3c/6818-1
00:34:26.752 --> 00:34:32.328
remember why I brought it and
anyway so that's that I I'll

025a3067-1e27-4a28-af18-1f93272d4d3c/6818-2
00:34:32.328 --> 00:34:34.029
show you the demo.

025a3067-1e27-4a28-af18-1f93272d4d3c/6832-0
00:34:34.039 --> 00:34:37.122
You know, after the the seminar
and the way it works in my

025a3067-1e27-4a28-af18-1f93272d4d3c/6832-1
00:34:37.122 --> 00:34:37.539
account.

025a3067-1e27-4a28-af18-1f93272d4d3c/6858-0
00:34:37.619 --> 00:34:40.996
Umm, yeah, so this is what I
have to say and I will be happy

025a3067-1e27-4a28-af18-1f93272d4d3c/6858-1
00:34:40.996 --> 00:34:42.379
to answer your questions.

025a3067-1e27-4a28-af18-1f93272d4d3c/6866-0
00:34:42.389 --> 00:34:45.409
And yeah, for I'll show you the
demo after that.

025a3067-1e27-4a28-af18-1f93272d4d3c/6884-0
00:34:52.629 --> 00:34:56.354
That switch audio from the show
computer over to the following

025a3067-1e27-4a28-af18-1f93272d4d3c/6884-1
00:34:56.354 --> 00:34:56.649
here.

025a3067-1e27-4a28-af18-1f93272d4d3c/6890-0
00:35:11.599 --> 00:35:12.109
OK.

025a3067-1e27-4a28-af18-1f93272d4d3c/6897-0
00:35:12.119 --> 00:35:14.069
Can folks in the audience here
Moffett.

025a3067-1e27-4a28-af18-1f93272d4d3c/6917-0
00:35:15.029 --> 00:35:18.144
Well, actually, you know what I
just asked you like you don't

025a3067-1e27-4a28-af18-1f93272d4d3c/6907-0
00:35:15.839 --> 00:35:15.999
Yes.

025a3067-1e27-4a28-af18-1f93272d4d3c/6917-1
00:35:18.144 --> 00:35:19.299
show a demo real quick.

025a3067-1e27-4a28-af18-1f93272d4d3c/6922-0
00:35:19.309 --> 00:35:20.299
I just have this sure.

025a3067-1e27-4a28-af18-1f93272d4d3c/6930-0
00:35:22.659 --> 00:35:23.849
We have the top again after.

025a3067-1e27-4a28-af18-1f93272d4d3c/6941-0
00:35:25.889 --> 00:35:27.659
You still can see my screen,
right?

025a3067-1e27-4a28-af18-1f93272d4d3c/6947-0
00:35:27.699 --> 00:35:30.429
Yeah, they're planning.

025a3067-1e27-4a28-af18-1f93272d4d3c/6950-0
00:35:31.929 --> 00:35:32.339
I just.

025a3067-1e27-4a28-af18-1f93272d4d3c/6989-0
00:35:32.379 --> 00:35:36.604
I just forgot to make it public
and OK, so face hunter you have,

025a3067-1e27-4a28-af18-1f93272d4d3c/6989-1
00:35:36.604 --> 00:35:40.699
you know, all the demo like you
saw it and now you can talk to

025a3067-1e27-4a28-af18-1f93272d4d3c/6989-2
00:35:40.699 --> 00:35:41.479
face Hunter.

025a3067-1e27-4a28-af18-1f93272d4d3c/7002-0
00:35:42.399 --> 00:35:46.929
And so the way you talk, you
will be like, right?

025a3067-1e27-4a28-af18-1f93272d4d3c/7008-0
00:35:46.939 --> 00:35:48.079
You never see my code.

025a3067-1e27-4a28-af18-1f93272d4d3c/7021-0
00:35:48.089 --> 00:35:50.549
Like you never interacted with
it and you would be interesting.

025a3067-1e27-4a28-af18-1f93272d4d3c/7051-0
00:35:51.019 --> 00:36:03.901
Hi how can I run facehunter on
Earth weeks from Moffett Field

025a3067-1e27-4a28-af18-1f93272d4d3c/7051-1
00:36:03.901 --> 00:36:08.679
on its of November and.

025a3067-1e27-4a28-af18-1f93272d4d3c/7066-0
00:36:11.459 --> 00:36:14.109
It says hello there are, you
know, obtaining seismic data.

025a3067-1e27-4a28-af18-1f93272d4d3c/7085-0
00:36:14.199 --> 00:36:18.231
Preprocess the data install face
Hunter, configure blah blah blah

025a3067-1e27-4a28-af18-1f93272d4d3c/7085-1
00:36:18.231 --> 00:36:20.429
blah blah blah blah blah blah
blah.

025a3067-1e27-4a28-af18-1f93272d4d3c/7090-0
00:36:21.869 --> 00:36:24.009
Show me info please.

025a3067-1e27-4a28-af18-1f93272d4d3c/7102-0
00:36:27.839 --> 00:36:30.919
Adidas right to the code to do
that?

025a3067-1e27-4a28-af18-1f93272d4d3c/7119-0
00:36:31.659 --> 00:36:36.619
Or even and it's still based on
like random ********.

025a3067-1e27-4a28-af18-1f93272d4d3c/7131-0
00:36:36.669 --> 00:36:40.679
It's based on my own guitar,
like it reads the repository.

025a3067-1e27-4a28-af18-1f93272d4d3c/7145-0
00:36:40.689 --> 00:36:43.749
Sorry for the word, but it reads
the repository. Resort.

025a3067-1e27-4a28-af18-1f93272d4d3c/7175-0
00:36:45.159 --> 00:36:48.794
Yeah, it reads the repository
and it actually goes through the

025a3067-1e27-4a28-af18-1f93272d4d3c/7175-1
00:36:48.794 --> 00:36:52.255
repository and like it sets up
the parameters and knows the

025a3067-1e27-4a28-af18-1f93272d4d3c/7175-2
00:36:52.255 --> 00:36:52.889
code there.

025a3067-1e27-4a28-af18-1f93272d4d3c/7202-0
00:36:52.939 --> 00:36:59.737
You can talk about the code you
you could be like umm, by the

025a3067-1e27-4a28-af18-1f93272d4d3c/7202-1
00:36:59.737 --> 00:37:04.889
way, why regression or let's see
feet cassion.

025a3067-1e27-4a28-af18-1f93272d4d3c/7231-0
00:37:11.079 --> 00:37:14.380
And you can literally talk to a
GitHub grapple and you can also

025a3067-1e27-4a28-af18-1f93272d4d3c/7231-1
00:37:14.380 --> 00:37:17.009
upload the paper in there, and
so on and so forth.

025a3067-1e27-4a28-af18-1f93272d4d3c/7252-0
00:37:17.079 --> 00:37:21.433
And this is coming, you know to
all of us with this new open AI

025a3067-1e27-4a28-af18-1f93272d4d3c/7252-1
00:37:21.433 --> 00:37:21.909
agents.

025a3067-1e27-4a28-af18-1f93272d4d3c/7265-0
00:37:21.979 --> 00:37:24.269
This is like what the whole fox
is gonna be about.

025a3067-1e27-4a28-af18-1f93272d4d3c/7279-0
00:37:24.419 --> 00:37:27.714
Like people are still trying to
understand what's going on

025a3067-1e27-4a28-af18-1f93272d4d3c/7279-1
00:37:27.714 --> 00:37:28.049
there.

025a3067-1e27-4a28-af18-1f93272d4d3c/7311-0
00:37:28.619 --> 00:37:32.444
So yeah, this is gonna be a lot
of fun, and it's gonna be very

025a3067-1e27-4a28-af18-1f93272d4d3c/7311-1
00:37:32.444 --> 00:37:36.026
transformative for the way we
share science for the way we

025a3067-1e27-4a28-af18-1f93272d4d3c/7311-2
00:37:36.026 --> 00:37:37.179
share our research.

025a3067-1e27-4a28-af18-1f93272d4d3c/7331-0
00:37:37.809 --> 00:37:41.210
And this like basically like a
new website like instead of

025a3067-1e27-4a28-af18-1f93272d4d3c/7331-1
00:37:41.210 --> 00:37:41.959
website yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/7333-0
00:37:41.969 --> 00:37:42.439
Now I'm done.

025a3067-1e27-4a28-af18-1f93272d4d3c/7335-0
00:37:42.449 --> 00:37:42.779
Thank you.

025a3067-1e27-4a28-af18-1f93272d4d3c/7342-0
00:37:47.719 --> 00:37:49.239
Questions online are in the
room.

025a3067-1e27-4a28-af18-1f93272d4d3c/7347-0
00:37:56.939 --> 00:37:57.059
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/7351-0
00:37:57.999 --> 00:37:58.569
Umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/7384-0
00:37:58.959 --> 00:38:02.249
So if if you were given access
to an agent, what would you use

025a3067-1e27-4a28-af18-1f93272d4d3c/7384-1
00:38:02.249 --> 00:38:05.331
your agent to do in terms of
science that you're not doing

025a3067-1e27-4a28-af18-1f93272d4d3c/7384-2
00:38:05.331 --> 00:38:05.539
now?

025a3067-1e27-4a28-af18-1f93272d4d3c/7390-0
00:38:06.699 --> 00:38:08.329
I guess it depends.

025a3067-1e27-4a28-af18-1f93272d4d3c/7405-0
00:38:08.739 --> 00:38:11.247
I don't think it looks like
really defined when agents are

025a3067-1e27-4a28-af18-1f93272d4d3c/7405-1
00:38:11.247 --> 00:38:11.799
for everyone.

025a3067-1e27-4a28-af18-1f93272d4d3c/7407-0
00:38:11.839 --> 00:38:12.119
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/7419-0
00:38:12.129 --> 00:38:13.799
So the agent is like what I
showed you.

025a3067-1e27-4a28-af18-1f93272d4d3c/7447-0
00:38:13.809 --> 00:38:19.402
The agent is like an instance of
a an AI that is able to access

025a3067-1e27-4a28-af18-1f93272d4d3c/7447-1
00:38:19.402 --> 00:38:23.159
certain resources and able to
act on that.

025a3067-1e27-4a28-af18-1f93272d4d3c/7468-0
00:38:23.729 --> 00:38:26.942
Right now, those agents have
very simple they're mostly

025a3067-1e27-4a28-af18-1f93272d4d3c/7468-1
00:38:26.942 --> 00:38:29.179
talking to agents like you don't
like.

025a3067-1e27-4a28-af18-1f93272d4d3c/7486-0
00:38:29.189 --> 00:38:33.332
They don't really use tools or
functions or anything but the

025a3067-1e27-4a28-af18-1f93272d4d3c/7486-1
00:38:33.332 --> 00:38:35.369
app for that is already there.

025a3067-1e27-4a28-af18-1f93272d4d3c/7520-0
00:38:35.379 --> 00:38:39.572
It's already published, so just
a matter of like days when we're

025a3067-1e27-4a28-af18-1f93272d4d3c/7520-1
00:38:39.572 --> 00:38:43.506
gonna see fancy edges and one of
the early examples of that,

025a3067-1e27-4a28-af18-1f93272d4d3c/7520-2
00:38:43.506 --> 00:38:44.989
someone took chat, GPT.

025a3067-1e27-4a28-af18-1f93272d4d3c/7553-0
00:38:44.999 --> 00:38:49.363
Now it has vision and they took
this chat GV division and they

025a3067-1e27-4a28-af18-1f93272d4d3c/7553-1
00:38:49.363 --> 00:38:53.311
ran a football game frame by
frame through chat, GPT and

025a3067-1e27-4a28-af18-1f93272d4d3c/7553-2
00:38:53.311 --> 00:38:55.319
asked it to commend the game.

025a3067-1e27-4a28-af18-1f93272d4d3c/7560-0
00:38:55.649 --> 00:38:56.579
And it does.

025a3067-1e27-4a28-af18-1f93272d4d3c/7581-0
00:38:56.649 --> 00:39:00.296
And it then the whisted with the
voice, because now it has the

025a3067-1e27-4a28-af18-1f93272d4d3c/7581-1
00:39:00.296 --> 00:39:01.859
voice you can talk to chat.

025a3067-1e27-4a28-af18-1f93272d4d3c/7620-0
00:39:01.869 --> 00:39:04.882
GPD like in Excel voice and it's
like you're just watching the

025a3067-1e27-4a28-af18-1f93272d4d3c/7620-1
00:39:04.882 --> 00:39:07.800
translation of football, you
know, and this is like, this is

025a3067-1e27-4a28-af18-1f93272d4d3c/7620-2
00:39:07.800 --> 00:39:09.569
happening the day after the
release.

025a3067-1e27-4a28-af18-1f93272d4d3c/7633-0
00:39:09.619 --> 00:39:11.099
So imagine what we're gonna have
in the week.

025a3067-1e27-4a28-af18-1f93272d4d3c/7640-0
00:39:11.109 --> 00:39:12.199
We're gonna have in the months.

025a3067-1e27-4a28-af18-1f93272d4d3c/7656-0
00:39:12.209 --> 00:39:15.344
It's like it's gonna be ask, you
know, as profound of an input as

025a3067-1e27-4a28-af18-1f93272d4d3c/7656-1
00:39:15.344 --> 00:39:16.199
changing yourself.

025a3067-1e27-4a28-af18-1f93272d4d3c/7675-0
00:39:16.249 --> 00:39:19.588
But Speaking of agents, so right
now I think they're great for

025a3067-1e27-4a28-af18-1f93272d4d3c/7675-1
00:39:19.588 --> 00:39:20.329
communicating.

025a3067-1e27-4a28-af18-1f93272d4d3c/7722-0
00:39:20.779 --> 00:39:24.647
I think they're great for
explaining your code and things

025a3067-1e27-4a28-af18-1f93272d4d3c/7722-1
00:39:24.647 --> 00:39:28.581
like this because it's much
easier to ask a chatpad how to

025a3067-1e27-4a28-af18-1f93272d4d3c/7722-2
00:39:28.581 --> 00:39:32.516
set up a face hunter for an
earthquake few days ago rather

025a3067-1e27-4a28-af18-1f93272d4d3c/7722-3
00:39:32.516 --> 00:39:34.849
than go through the code
yourself.

025a3067-1e27-4a28-af18-1f93272d4d3c/7734-0
00:39:34.899 --> 00:39:36.209
Even though I tried to make it
accessible.

025a3067-1e27-4a28-af18-1f93272d4d3c/7743-0
00:39:36.939 --> 00:39:39.769
Umm, what is gonna be for
research?

025a3067-1e27-4a28-af18-1f93272d4d3c/7752-0
00:39:39.779 --> 00:39:40.799
I don't even know.

025a3067-1e27-4a28-af18-1f93272d4d3c/7762-0
00:39:40.809 --> 00:39:43.609
My thing is, I think it's just
gonna be great.

025a3067-1e27-4a28-af18-1f93272d4d3c/7771-0
00:39:43.619 --> 00:39:46.229
Like, we're just gonna have,
like, brainstorm partners.

025a3067-1e27-4a28-af18-1f93272d4d3c/7805-0
00:39:46.269 --> 00:39:49.458
Like I'm you know, you know the
way you use tragedy now is gonna

025a3067-1e27-4a28-af18-1f93272d4d3c/7805-1
00:39:49.458 --> 00:39:52.402
be just amplified because you
could be you would be able to

025a3067-1e27-4a28-af18-1f93272d4d3c/7805-2
00:39:52.402 --> 00:39:53.039
customize it.

025a3067-1e27-4a28-af18-1f93272d4d3c/7828-0
00:39:53.379 --> 00:39:57.209
And I don't know about federal
government policies, but I use

025a3067-1e27-4a28-af18-1f93272d4d3c/7828-1
00:39:57.209 --> 00:40:00.359
tragic a lot, like, way more
than I care to admit.

025a3067-1e27-4a28-af18-1f93272d4d3c/7830-0
00:40:03.009 --> 00:40:03.209
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/7853-0
00:40:06.539 --> 00:40:11.862
Any questions that what are two
limitations of deep learning

025a3067-1e27-4a28-af18-1f93272d4d3c/7853-1
00:40:11.862 --> 00:40:16.399
over the traditional methods for
seismic expecting?

025a3067-1e27-4a28-af18-1f93272d4d3c/7860-0
00:40:17.159 --> 00:40:18.309
I don't think they are like.

025a3067-1e27-4a28-af18-1f93272d4d3c/7891-0
00:40:20.109 --> 00:40:23.577
There might be limitations of
any particular method in deep

025a3067-1e27-4a28-af18-1f93272d4d3c/7891-1
00:40:23.577 --> 00:40:27.160
learning, like what particular
architecture or particular you

025a3067-1e27-4a28-af18-1f93272d4d3c/7891-2
00:40:27.160 --> 00:40:29.529
know type of deep learning or
something.

025a3067-1e27-4a28-af18-1f93272d4d3c/7904-0
00:40:29.539 --> 00:40:31.606
But I don't think they're
limitations to deep learning

025a3067-1e27-4a28-af18-1f93272d4d3c/7904-1
00:40:31.606 --> 00:40:31.869
things.

025a3067-1e27-4a28-af18-1f93272d4d3c/7922-0
00:40:31.929 --> 00:40:35.160
It's it's just like what the
limitations, I mean, what does

025a3067-1e27-4a28-af18-1f93272d4d3c/7922-1
00:40:35.160 --> 00:40:36.829
the limitations of electricity?

025a3067-1e27-4a28-af18-1f93272d4d3c/7930-0
00:40:36.839 --> 00:40:39.779
What are they like in this
sense?

025a3067-1e27-4a28-af18-1f93272d4d3c/7948-0
00:40:40.089 --> 00:40:42.779
I obviously there are
limitations in this particular

025a3067-1e27-4a28-af18-1f93272d4d3c/7948-1
00:40:42.779 --> 00:40:44.199
instance, like if we have a.

025a3067-1e27-4a28-af18-1f93272d4d3c/7955-0
00:40:45.849 --> 00:40:46.539
Cattle.

025a3067-1e27-4a28-af18-1f93272d4d3c/7973-0
00:40:46.609 --> 00:40:49.622
There was invitations to that,
but you know that like this seem

025a3067-1e27-4a28-af18-1f93272d4d3c/7973-1
00:40:49.622 --> 00:40:51.269
it's like, do you see my
metaphor?

025a3067-1e27-4a28-af18-1f93272d4d3c/7987-0
00:40:51.319 --> 00:40:54.395
But I I'm comparing the
traditional methods with a deep

025a3067-1e27-4a28-af18-1f93272d4d3c/7987-1
00:40:54.395 --> 00:40:54.889
learning.

025a3067-1e27-4a28-af18-1f93272d4d3c/7998-0
00:40:54.999 --> 00:40:59.269
Yeah, I and let's say particular
the best implementation.

025a3067-1e27-4a28-af18-1f93272d4d3c/8047-0
00:40:59.279 --> 00:41:02.834
So one of the biggest deals
until recently was that you

025a3067-1e27-4a28-af18-1f93272d4d3c/8047-1
00:41:02.834 --> 00:41:07.024
never know if your prediction is
rubbish in deep learning and you

025a3067-1e27-4a28-af18-1f93272d4d3c/8047-2
00:41:07.024 --> 00:41:10.579
usually do know that with
classical traditional methods

025a3067-1e27-4a28-af18-1f93272d4d3c/8047-3
00:41:10.579 --> 00:41:13.689
based on hard owned physics and
so on the horse.

025a3067-1e27-4a28-af18-1f93272d4d3c/8069-0
00:41:13.769 --> 00:41:17.192
Now it's also changing, so he
said like, I guess it's like the

025a3067-1e27-4a28-af18-1f93272d4d3c/8069-1
00:41:17.192 --> 00:41:19.419
limitations have been pushed
away a bit.

025a3067-1e27-4a28-af18-1f93272d4d3c/8124-0
00:41:19.589 --> 00:41:23.158
Then it depends on any
particular task you're trying to

025a3067-1e27-4a28-af18-1f93272d4d3c/8124-1
00:41:23.158 --> 00:41:27.110
do, and for me, like for face
hunter, the limitation could be

025a3067-1e27-4a28-af18-1f93272d4d3c/8124-2
00:41:27.110 --> 00:41:30.870
the resources you ran it on,
like the the the CPUs or GPUs

025a3067-1e27-4a28-af18-1f93272d4d3c/8124-3
00:41:30.870 --> 00:41:34.757
that you're using for it, and
the computational resources to

025a3067-1e27-4a28-af18-1f93272d4d3c/8124-4
00:41:34.757 --> 00:41:35.139
train.

025a3067-1e27-4a28-af18-1f93272d4d3c/8150-0
00:41:35.229 --> 00:41:39.472
But to be fair, for particularly
in particular, face 100 doesn't

025a3067-1e27-4a28-af18-1f93272d4d3c/8150-1
00:41:39.472 --> 00:41:42.279
really take that much resource
into train.

025a3067-1e27-4a28-af18-1f93272d4d3c/8188-0
00:41:42.289 --> 00:41:47.340
You can train face Hunter on the
single GPU in like less than a

025a3067-1e27-4a28-af18-1f93272d4d3c/8188-1
00:41:47.340 --> 00:41:52.232
date on the data set of size
stat and even bigger and so it's

025a3067-1e27-4a28-af18-1f93272d4d3c/8188-2
00:41:52.232 --> 00:41:53.179
fast enough.

025a3067-1e27-4a28-af18-1f93272d4d3c/8221-0
00:41:53.189 --> 00:41:56.667
And also the advancements in
deep learning are happening with

025a3067-1e27-4a28-af18-1f93272d4d3c/8221-1
00:41:56.667 --> 00:42:00.089
such a staggering speed that
whatever limitations we have to

025a3067-1e27-4a28-af18-1f93272d4d3c/8221-2
00:42:00.089 --> 00:42:02.669
today, probably not gonna be a
deal tomorrow.

025a3067-1e27-4a28-af18-1f93272d4d3c/8257-0
00:42:03.219 --> 00:42:07.143
And because every like the deep
learning is the fastest growing

025a3067-1e27-4a28-af18-1f93272d4d3c/8257-1
00:42:07.143 --> 00:42:10.514
field in the world right now,
like every day is like a

025a3067-1e27-4a28-af18-1f93272d4d3c/8257-2
00:42:10.514 --> 00:42:12.169
hundreds of quality papers.

025a3067-1e27-4a28-af18-1f93272d4d3c/8267-0
00:42:12.359 --> 00:42:14.599
And it's impossible to give up
anymore.

025a3067-1e27-4a28-af18-1f93272d4d3c/8273-0
00:42:16.009 --> 00:42:17.069
And I'm really trying to.

025a3067-1e27-4a28-af18-1f93272d4d3c/8281-0
00:42:20.469 --> 00:42:22.239
Umm, I had a question.

025a3067-1e27-4a28-af18-1f93272d4d3c/8324-0
00:42:22.469 --> 00:42:26.669
You know, I'm somewhat of a
novice and seismology so maybe

025a3067-1e27-4a28-af18-1f93272d4d3c/8324-1
00:42:26.669 --> 00:42:31.367
this is an obvious question, but
you seem surprised that your uh,

025a3067-1e27-4a28-af18-1f93272d4d3c/8324-2
00:42:31.367 --> 00:42:35.496
you're approaching is able to
approximate depth with some

025a3067-1e27-4a28-af18-1f93272d4d3c/8324-3
00:42:35.496 --> 00:42:36.919
reasonable accuracy.

025a3067-1e27-4a28-af18-1f93272d4d3c/8337-0
00:42:37.289 --> 00:42:40.999
What within a seismogram might
lead to?

025a3067-1e27-4a28-af18-1f93272d4d3c/8348-0
00:42:41.059 --> 00:42:43.359
Like how can you predict depth
from a from a?

025a3067-1e27-4a28-af18-1f93272d4d3c/8353-0
00:42:43.409 --> 00:42:44.089
A seismic room.

025a3067-1e27-4a28-af18-1f93272d4d3c/8360-0
00:42:44.139 --> 00:42:45.079
Don't I have no idea.

025a3067-1e27-4a28-af18-1f93272d4d3c/8397-0
00:42:45.429 --> 00:42:49.602
So there are some papers showing
that, and if you just have the

025a3067-1e27-4a28-af18-1f93272d4d3c/8397-1
00:42:49.602 --> 00:42:53.579
seismograms and you can run them
like through some empirical

025a3067-1e27-4a28-af18-1f93272d4d3c/8397-2
00:42:53.579 --> 00:42:56.839
formula is something you can
estimate the depths.

025a3067-1e27-4a28-af18-1f93272d4d3c/8420-0
00:42:56.849 --> 00:43:00.229
I think the way it works, it
doesn't give running setting is

025a3067-1e27-4a28-af18-1f93272d4d3c/8420-1
00:43:00.229 --> 00:43:01.669
that you're probably like.

025a3067-1e27-4a28-af18-1f93272d4d3c/8513-0
00:43:01.679 --> 00:43:05.374
You're frequencies are getting
depleted in some particular way,

025a3067-1e27-4a28-af18-1f93272d4d3c/8513-1
00:43:05.374 --> 00:43:09.011
but the the reason why I was so
surprised is is because that's

025a3067-1e27-4a28-af18-1f93272d4d3c/8513-2
00:43:09.011 --> 00:43:12.186
is a very relative thing like
you can have a different

025a3067-1e27-4a28-af18-1f93272d4d3c/8513-3
00:43:12.186 --> 00:43:15.765
distance to their speaking the
different depths and depending

025a3067-1e27-4a28-af18-1f93272d4d3c/8513-4
00:43:15.765 --> 00:43:19.575
on the two you know you have the
different configurations and yet

025a3067-1e27-4a28-af18-1f93272d4d3c/8513-5
00:43:19.575 --> 00:43:23.155
like I tried to predict like
some weird way pass or something

025a3067-1e27-4a28-af18-1f93272d4d3c/8513-6
00:43:23.155 --> 00:43:26.734
and I'm not getting it right
like I'm not predicting it right

025a3067-1e27-4a28-af18-1f93272d4d3c/8513-7
00:43:26.734 --> 00:43:29.389
because I I'm not computing it
right I guess.

025a3067-1e27-4a28-af18-1f93272d4d3c/8535-0
00:43:29.459 --> 00:43:34.789
But then like somehow is able to
see different distance depths

025a3067-1e27-4a28-af18-1f93272d4d3c/8535-1
00:43:34.789 --> 00:43:37.919
without really knowing the
distance.

025a3067-1e27-4a28-af18-1f93272d4d3c/8573-0
00:43:38.009 --> 00:43:41.079
Like when I'm predicting the
distance it's it's independent

025a3067-1e27-4a28-af18-1f93272d4d3c/8573-1
00:43:41.079 --> 00:43:43.994
like I you know, even if you
have like 4 hats and you're

025a3067-1e27-4a28-af18-1f93272d4d3c/8573-2
00:43:43.994 --> 00:43:47.166
predicting 4 parameters, all of
them are kinda independent in

025a3067-1e27-4a28-af18-1f93272d4d3c/8573-3
00:43:47.166 --> 00:43:47.779
predictions.

025a3067-1e27-4a28-af18-1f93272d4d3c/8639-0
00:43:47.789 --> 00:43:51.085
So you don't really need one on
the other, so that's why I'm

025a3067-1e27-4a28-af18-1f93272d4d3c/8639-1
00:43:51.085 --> 00:43:54.435
very surprised about the devs,
not the fact that it's able to

025a3067-1e27-4a28-af18-1f93272d4d3c/8639-2
00:43:54.435 --> 00:43:57.947
do that, but the fact that it's
kind of like weirdly formulated,

025a3067-1e27-4a28-af18-1f93272d4d3c/8639-3
00:43:57.947 --> 00:44:01.081
like what is like, why why
depths like, well, why does it

025a3067-1e27-4a28-af18-1f93272d4d3c/8639-4
00:44:01.081 --> 00:44:02.269
work without distance?

025a3067-1e27-4a28-af18-1f93272d4d3c/8645-0
00:44:02.269 --> 00:44:03.949
Hill, this is what surprised me.

025a3067-1e27-4a28-af18-1f93272d4d3c/8647-0
00:44:04.499 --> 00:44:04.969
Gotcha.

025a3067-1e27-4a28-af18-1f93272d4d3c/8651-0
00:44:04.979 --> 00:44:05.639
So so you don't.

025a3067-1e27-4a28-af18-1f93272d4d3c/8685-0
00:44:05.649 --> 00:44:09.501
There's not necessarily a a
easily identifiable and easily

025a3067-1e27-4a28-af18-1f93272d4d3c/8685-1
00:44:09.501 --> 00:44:13.744
identifiable physical mechanism
that you can point to within the

025a3067-1e27-4a28-af18-1f93272d4d3c/8685-2
00:44:13.744 --> 00:44:16.159
seismogram like a single
seismogram.

025a3067-1e27-4a28-af18-1f93272d4d3c/8737-0
00:44:16.169 --> 00:44:19.375
And so I probably like I guess,
I mean, I don't really have as

025a3067-1e27-4a28-af18-1f93272d4d3c/8737-1
00:44:19.375 --> 00:44:22.377
much experience as some other
small use in the room, but I

025a3067-1e27-4a28-af18-1f93272d4d3c/8737-2
00:44:22.377 --> 00:44:25.430
won't be able to tell just by
looking at the setting ground

025a3067-1e27-4a28-af18-1f93272d4d3c/8737-3
00:44:25.430 --> 00:44:26.549
from a single station.

025a3067-1e27-4a28-af18-1f93272d4d3c/8767-0
00:44:26.699 --> 00:44:29.305
And you know, to be honest, even
if I would have like 100 sizing

025a3067-1e27-4a28-af18-1f93272d4d3c/8767-1
00:44:29.305 --> 00:44:31.069
grams just by looking at it, I
wouldn't be.

025a3067-1e27-4a28-af18-1f93272d4d3c/8769-0
00:44:31.389 --> 00:44:32.229
I'll do this.

025a3067-1e27-4a28-af18-1f93272d4d3c/8778-0
00:44:32.319 --> 00:44:33.669
I would have the compute
something.

025a3067-1e27-4a28-af18-1f93272d4d3c/8785-0
00:44:33.679 --> 00:44:34.689
Sure, sure, sure, sure.

025a3067-1e27-4a28-af18-1f93272d4d3c/8787-0
00:44:35.059 --> 00:44:35.929
You're educated.

025a3067-1e27-4a28-af18-1f93272d4d3c/8813-0
00:44:35.939 --> 00:44:39.774
Guess it's possible that the
deeper the earthquake the the

025a3067-1e27-4a28-af18-1f93272d4d3c/8813-1
00:44:39.774 --> 00:44:43.219
less you have to go through
through the upper crust.

025a3067-1e27-4a28-af18-1f93272d4d3c/8824-0
00:44:43.229 --> 00:44:47.849
The attenuating upper crust, the
so the frequency content will be

025a3067-1e27-4a28-af18-1f93272d4d3c/8824-1
00:44:47.849 --> 00:44:48.549
different.

025a3067-1e27-4a28-af18-1f93272d4d3c/8826-0
00:44:49.459 --> 00:44:49.599
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/8829-0
00:44:50.179 --> 00:44:50.869
Are there?

025a3067-1e27-4a28-af18-1f93272d4d3c/8833-0
00:44:51.039 --> 00:44:51.549
I'm sorry.

025a3067-1e27-4a28-af18-1f93272d4d3c/8835-0
00:44:51.559 --> 00:44:52.609
I'm interrupting.

025a3067-1e27-4a28-af18-1f93272d4d3c/8841-0
00:44:52.899 --> 00:44:55.329
Go for what I do this.

025a3067-1e27-4a28-af18-1f93272d4d3c/8852-0
00:44:59.049 --> 00:45:00.999
There are and I'm not.

025a3067-1e27-4a28-af18-1f93272d4d3c/8933-0
00:45:01.249 --> 00:45:07.636
I I know very little about about
your your specialty field, but I

025a3067-1e27-4a28-af18-1f93272d4d3c/8933-1
00:45:07.636 --> 00:45:13.249
understand there is there is
intelligent AI that might be

025a3067-1e27-4a28-af18-1f93272d4d3c/8933-2
00:45:13.249 --> 00:45:19.153
able to address the question
that was asked and you can back

025a3067-1e27-4a28-af18-1f93272d4d3c/8933-3
00:45:19.153 --> 00:45:25.443
out a little bit why you got the
answer you right, uh why are we

025a3067-1e27-4a28-af18-1f93272d4d3c/8933-4
00:45:25.443 --> 00:45:31.346
able to print dict laps from a
single size small round 3 red

025a3067-1e27-4a28-af18-1f93272d4d3c/8933-5
00:45:31.346 --> 00:45:37.056
cards the list of the distance
no promises that it's gonna

025a3067-1e27-4a28-af18-1f93272d4d3c/8933-6
00:45:37.056 --> 00:45:37.539
work.

025a3067-1e27-4a28-af18-1f93272d4d3c/8939-0
00:45:39.339 --> 00:45:41.919
It's very, very positive.

025a3067-1e27-4a28-af18-1f93272d4d3c/8945-0
00:45:44.589 --> 00:45:44.959
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/8950-0
00:45:44.969 --> 00:45:46.229
Well, seismic wave radiates.

025a3067-1e27-4a28-af18-1f93272d4d3c/8953-0
00:45:46.359 --> 00:45:46.769
Thank you.

025a3067-1e27-4a28-af18-1f93272d4d3c/8955-0
00:45:46.899 --> 00:45:47.099
OK.

025a3067-1e27-4a28-af18-1f93272d4d3c/8965-0
00:45:48.359 --> 00:45:53.479
Hey, good point summary I have.

025a3067-1e27-4a28-af18-1f93272d4d3c/9011-0
00:45:53.529 --> 00:45:56.394
I have face Hunter again
utilizes techniques like

025a3067-1e27-4a28-af18-1f93272d4d3c/9011-1
00:45:56.394 --> 00:45:59.659
waveform analysis, pattern
recognition, say from grounds

025a3067-1e27-4a28-af18-1f93272d4d3c/9011-2
00:45:59.659 --> 00:46:03.039
records have been based rules
for the first layers from an

025a3067-1e27-4a28-af18-1f93272d4d3c/9011-3
00:46:03.039 --> 00:46:03.669
earthquake.

025a3067-1e27-4a28-af18-1f93272d4d3c/9015-0
00:46:03.679 --> 00:46:07.159
Source says phases such and.

025a3067-1e27-4a28-af18-1f93272d4d3c/9045-0
00:46:09.579 --> 00:46:12.834
No, I think like I think it's
stuck in the phases because this

025a3067-1e27-4a28-af18-1f93272d4d3c/9045-1
00:46:12.834 --> 00:46:15.778
is a particular ration version
that has the code for the

025a3067-1e27-4a28-af18-1f93272d4d3c/9045-2
00:46:15.778 --> 00:46:16.139
phases.

025a3067-1e27-4a28-af18-1f93272d4d3c/9067-0
00:46:16.149 --> 00:46:19.542
It doesn't have the code for
depths, but surely would be

025a3067-1e27-4a28-af18-1f93272d4d3c/9067-1
00:46:19.542 --> 00:46:22.219
interesting to ask if we replace
the depths.

025a3067-1e27-4a28-af18-1f93272d4d3c/9069-0
00:46:22.229 --> 00:46:22.479
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/9080-0
00:46:22.519 --> 00:46:24.689
Did you train it with the
labeled?

025a3067-1e27-4a28-af18-1f93272d4d3c/9092-0
00:46:25.259 --> 00:46:25.569
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/9105-0
00:46:25.579 --> 00:46:28.209
Like when I when I demonstrated
this examples of like one

025a3067-1e27-4a28-af18-1f93272d4d3c/9105-1
00:46:28.209 --> 00:46:29.569
second, two seconds 3 seconds.

025a3067-1e27-4a28-af18-1f93272d4d3c/9107-0
00:46:29.579 --> 00:46:29.779
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/9127-0
00:46:29.819 --> 00:46:32.057
I will, of course, it
demonstrated it with devs

025a3067-1e27-4a28-af18-1f93272d4d3c/9127-1
00:46:32.057 --> 00:46:32.429
labeled.

025a3067-1e27-4a28-af18-1f93272d4d3c/9144-0
00:46:32.439 --> 00:46:36.389
OK, won't be able to like, you
know, like the agent I'm showing

025a3067-1e27-4a28-af18-1f93272d4d3c/9144-1
00:46:36.389 --> 00:46:37.129
here, right?

025a3067-1e27-4a28-af18-1f93272d4d3c/9157-0
00:46:37.139 --> 00:46:39.059
Like this is like an advanced
technology.

025a3067-1e27-4a28-af18-1f93272d4d3c/9170-0
00:46:39.509 --> 00:46:43.659
The deep learning method of face
hunter is very specialized tool.

025a3067-1e27-4a28-af18-1f93272d4d3c/9182-0
00:46:43.789 --> 00:46:46.339
It's like a A I mean, this is
what it is, right?

025a3067-1e27-4a28-af18-1f93272d4d3c/9194-0
00:46:46.349 --> 00:46:48.779
It's a tool that, you know, like
it has it purpose.

025a3067-1e27-4a28-af18-1f93272d4d3c/9224-0
00:46:48.849 --> 00:46:52.267
You can reshape the tool really
easily to do like to apply to

025a3067-1e27-4a28-af18-1f93272d4d3c/9224-1
00:46:52.267 --> 00:46:55.353
different purposes, which is
different from traditional

025a3067-1e27-4a28-af18-1f93272d4d3c/9224-2
00:46:55.353 --> 00:46:55.849
software.

025a3067-1e27-4a28-af18-1f93272d4d3c/9248-0
00:46:55.859 --> 00:47:00.151
That is like more region in a
sense, but there's still a tool

025a3067-1e27-4a28-af18-1f93272d4d3c/9248-1
00:47:00.151 --> 00:47:02.989
that doesn't really do anything
pensive.

025a3067-1e27-4a28-af18-1f93272d4d3c/9251-0
00:47:05.849 --> 00:47:06.299
Alright.

025a3067-1e27-4a28-af18-1f93272d4d3c/9266-0
00:47:07.329 --> 00:47:11.139
Yeah, I'm you guys can try to do
all of that yourself.

025a3067-1e27-4a28-af18-1f93272d4d3c/9292-0
00:47:11.149 --> 00:47:14.570
It's very easy, like it takes a
literal second to like OK,

025a3067-1e27-4a28-af18-1f93272d4d3c/9292-1
00:47:14.570 --> 00:47:16.599
minute to set up a booth
yourself.

025a3067-1e27-4a28-af18-1f93272d4d3c/9338-0
00:47:16.669 --> 00:47:20.226
There reason why I wasn't able
to show it in a in a in a

025a3067-1e27-4a28-af18-1f93272d4d3c/9338-1
00:47:20.226 --> 00:47:24.033
different account is because
like I have a developer preview

025a3067-1e27-4a28-af18-1f93272d4d3c/9338-2
00:47:24.033 --> 00:47:27.714
like new features that are
coming and these features allow

025a3067-1e27-4a28-af18-1f93272d4d3c/9338-3
00:47:27.714 --> 00:47:29.149
you to like upload the.

025a3067-1e27-4a28-af18-1f93272d4d3c/9361-0
00:47:31.719 --> 00:47:35.831
There's like a a field here that
you use to upload your your

025a3067-1e27-4a28-af18-1f93272d4d3c/9361-1
00:47:35.831 --> 00:47:37.179
files into the boat.

025a3067-1e27-4a28-af18-1f93272d4d3c/9363-0
00:47:37.979 --> 00:47:38.329
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/9365-0
00:47:38.379 --> 00:47:38.889
Alright.

025a3067-1e27-4a28-af18-1f93272d4d3c/9368-0
00:47:38.989 --> 00:47:39.629
Umm.

025a3067-1e27-4a28-af18-1f93272d4d3c/9387-0
00:47:40.139 --> 00:47:43.574
And keep an eye on the open this
eye agent store and it's gonna

025a3067-1e27-4a28-af18-1f93272d4d3c/9387-1
00:47:43.574 --> 00:47:44.539
be transformative.

025a3067-1e27-4a28-af18-1f93272d4d3c/9391-0
00:47:44.549 --> 00:47:45.129
Whatever one.

025a3067-1e27-4a28-af18-1f93272d4d3c/9398-0
00:47:45.529 --> 00:47:48.699
And so the I had of the curve.

025a3067-1e27-4a28-af18-1f93272d4d3c/9409-0
00:47:49.189 --> 00:47:52.019
Thank you so much for letting me
to speak here.

025a3067-1e27-4a28-af18-1f93272d4d3c/9415-0
00:47:52.029 --> 00:47:53.639
It's a great pleasure.

025a3067-1e27-4a28-af18-1f93272d4d3c/9418-0
00:47:54.469 --> 00:47:54.899
Yeah.

025a3067-1e27-4a28-af18-1f93272d4d3c/9436-0
00:47:54.949 --> 00:47:57.454
And I'm looking forward to, you
know, to having lively

025a3067-1e27-4a28-af18-1f93272d4d3c/9436-1
00:47:57.454 --> 00:47:58.319
discussion on this.

025a3067-1e27-4a28-af18-1f93272d4d3c/9438-0
00:47:58.869 --> 00:47:59.339
All right.

025a3067-1e27-4a28-af18-1f93272d4d3c/9445-0
00:47:59.349 --> 00:48:00.399
Let's thank our speaker again.