whisper-finetuning-for-asee / CS_441_2023_Spring_January_17,_2023.vtt
classen3's picture
Imported CS 441 audio/transcripts
a67be9a verified
WEBVTT Kind: captions; Language: en-US
NOTE
Created on 2024-02-07T20:43:56.2906687Z by ClassTranscribe
00:03:01.710 --> 00:03:04.290
Alright, good morning everybody.
00:03:06.080 --> 00:03:07.190
It's good to see you all.
00:03:07.190 --> 00:03:10.625
Hope you had a good break, so I'm going
00:03:10.625 --> 00:03:11.430
to get started.
00:03:12.600 --> 00:03:15.170
So this is CS441 Applied machine
00:03:15.170 --> 00:03:17.340
learning and I'm Derek Hoiem, the
00:03:17.340 --> 00:03:17.930
Instructor.
00:03:19.320 --> 00:03:21.240
So today I'm going to just tell you a
00:03:21.240 --> 00:03:22.510
little bit about myself.
00:03:22.510 --> 00:03:24.120
I'll talk about machine learning,
00:03:24.120 --> 00:03:27.030
applied machine learning and a bit
00:03:27.030 --> 00:03:28.530
about the course.
00:03:28.530 --> 00:03:30.616
I'll give an outline in the course and
00:03:30.616 --> 00:03:32.376
talk about some of the logistics of the
00:03:32.376 --> 00:03:32.579
course.
00:03:33.190 --> 00:03:34.510
And we'll probably end a little bit
00:03:34.510 --> 00:03:36.480
early so that there's time for you to
00:03:36.480 --> 00:03:37.840
ask any questions that you have about
00:03:37.840 --> 00:03:41.390
the course, either individually or you
00:03:41.390 --> 00:03:44.247
can ask them at the ask them in general
00:03:44.247 --> 00:03:45.880
at the end, at the end of class.
00:03:47.860 --> 00:03:50.440
So first a little bit about me.
00:03:50.440 --> 00:03:51.670
You might know some of this if you've
00:03:51.670 --> 00:03:53.890
taken a class with me before, but I was
00:03:53.890 --> 00:03:55.800
raised in upstate New York, right where
00:03:55.800 --> 00:03:57.240
the circle is and the Hudson River
00:03:57.240 --> 00:03:57.630
valley.
00:03:57.630 --> 00:03:59.320
This is the lake that I grew up on.
00:04:00.610 --> 00:04:03.730
I actually was interested in machine
00:04:03.730 --> 00:04:06.380
learning and AI for quite a long time.
00:04:07.750 --> 00:04:09.650
When I decided to go when I went to
00:04:09.650 --> 00:04:11.730
undergrad at Buffalo, I decided to
00:04:11.730 --> 00:04:14.280
major in electrical engineering because
00:04:14.280 --> 00:04:15.420
I.
00:04:15.820 --> 00:04:17.565
Because it has a good strong background
00:04:17.565 --> 00:04:20.070
in math, and I was advised that it's a
00:04:20.070 --> 00:04:21.540
good foundation for anything else I
00:04:21.540 --> 00:04:22.430
wanted to do.
00:04:22.430 --> 00:04:25.170
But my roommate was taking computer
00:04:25.170 --> 00:04:27.490
science and I liked his Assignments,
00:04:27.490 --> 00:04:29.050
and I started just doing them for fun
00:04:29.050 --> 00:04:31.610
and started taking more CS courses.
00:04:31.610 --> 00:04:33.360
And I kind of made a beeline in my
00:04:33.360 --> 00:04:36.239
curriculum for AI and machine learning,
00:04:36.240 --> 00:04:38.020
which we're not nearly as popular.
00:04:38.020 --> 00:04:39.830
At the time I was there was only one
00:04:39.830 --> 00:04:41.435
graduate course in machine learning,
00:04:41.435 --> 00:04:43.860
and I was the only undergraduate who
00:04:43.860 --> 00:04:45.940
took that course that year.
00:04:47.650 --> 00:04:50.828
And so I ended up adding on the
00:04:50.828 --> 00:04:52.880
computer the computer engineering
00:04:52.880 --> 00:04:54.240
degree as well, just because I had
00:04:54.240 --> 00:04:57.340
taken so many CS courses that I was
00:04:57.340 --> 00:04:58.890
just able to do that.
00:05:00.000 --> 00:05:02.010
When I after I graduated, I went to
00:05:02.010 --> 00:05:03.570
Carnegie Mellon and I went for
00:05:03.570 --> 00:05:04.410
Robotics.
00:05:04.600 --> 00:05:08.260
And I started working with somebody
00:05:08.260 --> 00:05:10.866
named Henry Schneiderman, who made at
00:05:10.866 --> 00:05:12.500
the what was at the time the most
00:05:12.500 --> 00:05:15.580
accurate face detector, and then.
00:05:15.660 --> 00:05:19.260
After he left to start a company and
00:05:19.260 --> 00:05:21.120
then I started working with Elisha
00:05:21.120 --> 00:05:24.810
Frozen Marshalli Bear doing applying
00:05:24.810 --> 00:05:26.770
machine learning to try to recognize
00:05:26.770 --> 00:05:29.430
geometry to create 3D models based on
00:05:29.430 --> 00:05:31.210
single view recognition.
00:05:32.140 --> 00:05:35.175
Then, after my PhD, I came here.
00:05:35.175 --> 00:05:38.280
I did like a postdoc fellowship for a
00:05:38.280 --> 00:05:39.360
little bit, and then I've been a
00:05:39.360 --> 00:05:40.720
professor here ever since.
00:05:43.140 --> 00:05:44.780
So I'll tell you a little bit about my
00:05:44.780 --> 00:05:45.530
research.
00:05:45.530 --> 00:05:47.870
Actually the first two projects that I
00:05:47.870 --> 00:05:51.480
did were on music identification and
00:05:51.480 --> 00:05:54.440
sound detection and then this is the.
00:05:54.440 --> 00:05:57.460
And then I did another one on object on
00:05:57.460 --> 00:05:59.820
retrieving images based on objects in
00:05:59.820 --> 00:06:00.050
them.
00:06:00.620 --> 00:06:02.440
And then this was like my first main
00:06:02.440 --> 00:06:04.970
project which was part of my thesis.
00:06:04.970 --> 00:06:08.310
So at the time, if you wanted to create
00:06:08.310 --> 00:06:10.760
a 3D model of a scene from images, you
00:06:10.760 --> 00:06:12.710
basically solved a bunch of algebraic
00:06:12.710 --> 00:06:13.460
constraints.
00:06:14.160 --> 00:06:15.860
Based on correspondences between the
00:06:15.860 --> 00:06:16.630
images.
00:06:16.630 --> 00:06:19.680
But we had this idea that since people
00:06:19.680 --> 00:06:22.010
are able to see an image like this one
00:06:22.010 --> 00:06:24.823
and understand the 3D scene behind the
00:06:24.823 --> 00:06:26.650
image, maybe we can use machine
00:06:26.650 --> 00:06:29.090
learning to recognize the surfaces and
00:06:29.090 --> 00:06:31.540
the image and then use that to create a
00:06:31.540 --> 00:06:32.650
simple 3D model.
00:06:35.370 --> 00:06:37.410
Currently a couple of the main
00:06:37.410 --> 00:06:40.625
directions I work in are one of them is
00:06:40.625 --> 00:06:42.070
this thing called neural radiance
00:06:42.070 --> 00:06:44.610
fields, and basically the idea is to
00:06:44.610 --> 00:06:47.360
use the machine learning system as a
00:06:47.360 --> 00:06:50.000
kind of compression to encode all the
00:06:50.000 --> 00:06:51.905
geometry and appearance information in
00:06:51.905 --> 00:06:54.525
the scene so that you can render out
00:06:54.525 --> 00:06:57.895
the surface normals or the depth or the
00:06:57.895 --> 00:06:59.940
appearance of images from arbitrary
00:06:59.940 --> 00:07:00.277
viewpoints.
00:07:00.277 --> 00:07:02.973
And so this is one of the directions
00:07:02.973 --> 00:07:03.830
that we're working.
00:07:04.500 --> 00:07:06.470
And another one is what I call general
00:07:06.470 --> 00:07:07.650
purpose Learning.
00:07:07.960 --> 00:07:10.930
And as you'll see, in a lot of machine
00:07:10.930 --> 00:07:13.190
learning, the goal is to solve one
00:07:13.190 --> 00:07:16.620
particular kind of Prediction task, or
00:07:16.620 --> 00:07:20.120
like unsupervised learning task.
00:07:20.120 --> 00:07:22.525
But people were quite differently.
00:07:22.525 --> 00:07:25.250
We don't have like a single classifier
00:07:25.250 --> 00:07:27.905
in our head or a single purpose that
00:07:27.905 --> 00:07:29.900
our mind is designed for.
00:07:29.900 --> 00:07:32.916
Instead, we have the ability to sense a
00:07:32.916 --> 00:07:34.638
lot of things and then we can do a lot
00:07:34.638 --> 00:07:36.650
of things with our body and our speech,
00:07:36.650 --> 00:07:38.530
and so we're able to solve kind of an
00:07:38.530 --> 00:07:39.310
infinite range.
00:07:39.400 --> 00:07:43.100
Tasks that are senses and our actions
00:07:43.100 --> 00:07:44.450
allow us to perform.
00:07:44.450 --> 00:07:46.140
And so we're trying to do the same
00:07:46.140 --> 00:07:47.790
thing for machine learning, to create
00:07:47.790 --> 00:07:50.493
systems that are able to receive some
00:07:50.493 --> 00:07:53.538
kinds of inputs, like text and images,
00:07:53.538 --> 00:07:56.330
and produce outputs which could also be
00:07:56.330 --> 00:07:57.457
like text or images.
00:07:57.457 --> 00:08:00.180
And then do any kinds of tasks that
00:08:00.180 --> 00:08:02.040
fall within that range of the input and
00:08:02.040 --> 00:08:02.833
output modality.
00:08:02.833 --> 00:08:05.240
And then we also work on ways to try to
00:08:05.240 --> 00:08:08.380
extend these systems, for example in
00:08:08.380 --> 00:08:08.750
this.
00:08:09.410 --> 00:08:11.260
And this slide here is showing that we
00:08:11.260 --> 00:08:14.000
created this system that is able to map
00:08:14.000 --> 00:08:16.420
from images in a text prompt into some
00:08:16.420 --> 00:08:17.320
kind of answer.
00:08:18.080 --> 00:08:20.260
And then the system is able to learn
00:08:20.260 --> 00:08:23.470
from web images to learn new concepts.
00:08:23.470 --> 00:08:25.550
So we were able to, we provided it with
00:08:25.550 --> 00:08:27.682
like a set of keywords about COVID, and
00:08:27.682 --> 00:08:29.760
then it was able to just download a
00:08:29.760 --> 00:08:31.863
bunch of images from Bing and then
00:08:31.863 --> 00:08:33.530
learn how to answer questions and
00:08:33.530 --> 00:08:35.300
caption and detect things that are
00:08:35.300 --> 00:08:37.290
related to COVID, which we're not in
00:08:37.290 --> 00:08:38.795
any of the original data that I trained
00:08:38.795 --> 00:08:39.050
in.
00:08:42.460 --> 00:08:44.460
I've also used machine learning in a
00:08:44.460 --> 00:08:47.414
lot of other ways, so let me just.
00:08:47.414 --> 00:08:49.045
I just want to make sure I think the
00:08:49.045 --> 00:08:50.650
mic's working, but I just want to.
00:08:51.720 --> 00:08:54.230
Make sure that it's OK.
00:08:54.230 --> 00:08:56.480
It'll pick up a little better down here
00:08:56.480 --> 00:08:58.370
because for the recording.
00:09:00.030 --> 00:09:02.375
So I've used machine learning in lots
00:09:02.375 --> 00:09:02.960
of different ways.
00:09:02.960 --> 00:09:05.540
In my research I've done things like
00:09:05.540 --> 00:09:07.515
object detection, image classification,
00:09:07.515 --> 00:09:10.020
3D scene modeling, Generating
00:09:10.020 --> 00:09:12.300
animations, visual question answering,
00:09:12.300 --> 00:09:14.240
phrase grounding, sound detection.
00:09:14.240 --> 00:09:16.690
So lots of different, lots of different
00:09:16.690 --> 00:09:17.390
use cases.
00:09:18.850 --> 00:09:20.566
And then I've also used machine
00:09:20.566 --> 00:09:22.700
learning in Application.
00:09:22.700 --> 00:09:25.240
So I cofounded this company,
00:09:25.240 --> 00:09:27.420
Reconstruct, where we take images of a
00:09:27.420 --> 00:09:29.460
construction site and bring it together
00:09:29.460 --> 00:09:31.525
with building plans and schedule to
00:09:31.525 --> 00:09:32.955
help all the stakeholders understand
00:09:32.955 --> 00:09:35.480
the progress and whether things are
00:09:35.480 --> 00:09:37.230
being built according to the plan.
00:09:37.230 --> 00:09:40.694
And part of this, some of the 3D
00:09:40.694 --> 00:09:42.020
construction doesn't use machine
00:09:42.020 --> 00:09:43.990
learning, but some of it does to help
00:09:43.990 --> 00:09:47.170
create more complete Models and then to
00:09:47.170 --> 00:09:48.880
recognize things in the scene.
00:09:48.950 --> 00:09:51.060
To help monitor their progress.
00:09:53.330 --> 00:09:56.100
So all of that is to say that I've had
00:09:56.100 --> 00:09:58.100
quite a lot of experience using machine
00:09:58.100 --> 00:09:59.380
learning in a variety of different
00:09:59.380 --> 00:10:02.205
ways, many different modalities, many
00:10:02.205 --> 00:10:04.940
different kinds of applications, both
00:10:04.940 --> 00:10:08.220
research and in commercial types of
00:10:08.220 --> 00:10:08.620
applications.
00:10:10.360 --> 00:10:12.267
So what is machine learning?
00:10:12.267 --> 00:10:14.170
So everyone has their own definition.
00:10:14.170 --> 00:10:16.740
But the way that I think about it is
00:10:16.740 --> 00:10:19.195
that it's machine learning spins Raw
00:10:19.195 --> 00:10:21.710
data into gold, so it's pretty easy to
00:10:21.710 --> 00:10:22.180
get data.
00:10:22.180 --> 00:10:23.880
Now there's all kinds of data.
00:10:23.880 --> 00:10:26.210
Whenever you use applications, you're
00:10:26.210 --> 00:10:29.500
providing those Application providers
00:10:29.500 --> 00:10:30.900
with your data.
00:10:30.900 --> 00:10:33.180
A lot of times you can download lots of
00:10:33.180 --> 00:10:34.470
information on the Internet.
00:10:34.470 --> 00:10:36.850
There's just data everywhere, but by
00:10:36.850 --> 00:10:38.970
itself, that data is not very useful.
00:10:39.030 --> 00:10:41.270
It's too much to manually go through.
00:10:41.270 --> 00:10:42.610
It's not something that you can
00:10:42.610 --> 00:10:46.320
actually solve a problem with, and so
00:10:46.320 --> 00:10:50.460
machine learning is really the ability
00:10:50.460 --> 00:10:52.610
to take all of that Raw data and turn
00:10:52.610 --> 00:10:55.066
it into something useful to be able to
00:10:55.066 --> 00:10:57.390
create predictive models or to create
00:10:57.390 --> 00:10:58.080
insights.
00:10:58.710 --> 00:11:01.760
So some examples are here for the Alexa
00:11:01.760 --> 00:11:04.060
speech recognition where you.
00:11:04.340 --> 00:11:06.940
Where they learn from audio
00:11:06.940 --> 00:11:09.920
transcriptions to be able to take the
00:11:09.920 --> 00:11:13.465
voice information that you speak into
00:11:13.465 --> 00:11:16.235
the Alexa app and turn it into text and
00:11:16.235 --> 00:11:18.390
then it's then turned into other
00:11:18.390 --> 00:11:19.310
information.
00:11:20.450 --> 00:11:23.110
Or product recommendations?
00:11:23.110 --> 00:11:24.860
Autonomous vehicles?
00:11:26.330 --> 00:11:30.186
Text generation like GPT 3 or GPT chat
00:11:30.186 --> 00:11:32.230
or ChatGPT image generation.
00:11:32.230 --> 00:11:34.675
And then there's a link there that you
00:11:34.675 --> 00:11:35.580
can check out later.
00:11:35.580 --> 00:11:37.780
It's a data visualization of Twitter
00:11:37.780 --> 00:11:40.780
where trying to take like all the tons
00:11:40.780 --> 00:11:42.400
of information about different tweets
00:11:42.400 --> 00:11:44.642
and organize it in a way so that you
00:11:44.642 --> 00:11:47.240
can draw insights about like social
00:11:47.240 --> 00:11:49.780
trends and kind of what's going on in
00:11:49.780 --> 00:11:50.630
different places.
00:11:54.780 --> 00:11:58.300
So we tend to think in classes.
00:11:58.300 --> 00:12:00.710
Especially we tend to think about
00:12:00.710 --> 00:12:03.250
machine learning as a problem of
00:12:03.250 --> 00:12:09.216
designing algorithms that will map your
00:12:09.216 --> 00:12:11.900
that allow you to learn from the
00:12:11.900 --> 00:12:14.410
training data and to achieve good test
00:12:14.410 --> 00:12:16.880
performance in your Prediction task
00:12:16.880 --> 00:12:18.150
according to the test data.
00:12:18.940 --> 00:12:20.920
But in the real world, there's actually
00:12:20.920 --> 00:12:22.310
a bigger problem.
00:12:22.310 --> 00:12:24.646
There's a bigger infrastructure around
00:12:24.646 --> 00:12:25.518
machine learning.
00:12:25.518 --> 00:12:27.945
And so while we're going to focus on
00:12:27.945 --> 00:12:30.240
the Algorithm and model development,
00:12:30.240 --> 00:12:32.099
it's also important to be aware of the
00:12:32.100 --> 00:12:34.154
broader context of machine learning.
00:12:34.154 --> 00:12:36.800
So if you were to ever go work for a
00:12:36.800 --> 00:12:39.465
company and they say that they want you
00:12:39.465 --> 00:12:43.100
to build some kind of predictor or some
00:12:43.100 --> 00:12:45.583
data cluster or something like that,
00:12:45.583 --> 00:12:48.230
you would actually go through multiple
00:12:48.230 --> 00:12:49.590
different phases, so the 1st.
00:12:50.360 --> 00:12:52.445
And potentially the most important is
00:12:52.445 --> 00:12:54.710
the data preparation where you collect
00:12:54.710 --> 00:12:56.326
and curate the data.
00:12:56.326 --> 00:12:59.147
So you have to find examples you have.
00:12:59.147 --> 00:13:00.970
You may need to Annotate it or hire
00:13:00.970 --> 00:13:03.370
annotators or create annotation tools.
00:13:03.370 --> 00:13:05.858
And then you split your data often into
00:13:05.858 --> 00:13:07.740
a training set, validation set and a
00:13:07.740 --> 00:13:08.278
test set.
00:13:08.278 --> 00:13:10.350
So the training set is what you would
00:13:10.350 --> 00:13:11.938
use to tune your Models, the validation
00:13:11.938 --> 00:13:14.441
is what you use to select your Models,
00:13:14.441 --> 00:13:17.685
and the test set is for your final
00:13:17.685 --> 00:13:20.210
Final like estimate of your systems.
00:13:20.270 --> 00:13:21.060
Performance.
00:13:22.370 --> 00:13:23.920
Then once you have that, then you can
00:13:23.920 --> 00:13:25.115
develop your algorithm.
00:13:25.115 --> 00:13:27.560
You can decide what kind of machine
00:13:27.560 --> 00:13:29.010
learning algorithm you're going to use.
00:13:29.010 --> 00:13:30.560
What are the hyperparameters.
00:13:31.570 --> 00:13:33.520
What kinds of objectives you're going
00:13:33.520 --> 00:13:37.120
to give to your Algorithm and you train
00:13:37.120 --> 00:13:38.850
it and evaluate it?
00:13:38.850 --> 00:13:41.350
Often that process takes quite a lot of
00:13:41.350 --> 00:13:44.193
time and it may lead you to go back to
00:13:44.193 --> 00:13:46.050
the data preparation stage if you
00:13:46.050 --> 00:13:47.400
realize that you're not going to be
00:13:47.400 --> 00:13:49.960
able to reach the applications
00:13:49.960 --> 00:13:51.030
performance requirements.
00:13:52.350 --> 00:13:54.100
Once you're finally feel like you've
00:13:54.100 --> 00:13:55.922
finished developing the Algorithm, then
00:13:55.922 --> 00:13:58.430
you would test it on a test set, which
00:13:58.430 --> 00:14:00.230
ideally you haven't used at any stage
00:14:00.230 --> 00:14:01.520
before, so that it gives you an
00:14:01.520 --> 00:14:03.480
unbiased estimate of your performance,
00:14:03.480 --> 00:14:05.640
and then finally you integrate it into
00:14:05.640 --> 00:14:06.570
your Application.
00:14:07.210 --> 00:14:08.600
And.
00:14:08.710 --> 00:14:11.160
And usually this Prediction engine that
00:14:11.160 --> 00:14:13.170
you've built will only be one part of
00:14:13.170 --> 00:14:15.410
an entire solution, and so it needs to
00:14:15.410 --> 00:14:16.853
work well with the rest of that
00:14:16.853 --> 00:14:17.139
solution.
00:14:18.420 --> 00:14:22.480
So as an example, consider the voice
00:14:22.480 --> 00:14:24.685
recognition and Alexa.
00:14:24.685 --> 00:14:27.896
So they had a really challenging
00:14:27.896 --> 00:14:30.540
problem that probably many of you have
00:14:30.540 --> 00:14:34.120
at least used Alexa app sometime.
00:14:34.120 --> 00:14:35.696
But there's a really challenging
00:14:35.696 --> 00:14:37.370
problem that you've got like some disc
00:14:37.370 --> 00:14:40.280
with some microphones on it and it you
00:14:40.280 --> 00:14:43.945
need the disk needs to be able to
00:14:43.945 --> 00:14:45.855
interpret your speech and it needs to
00:14:45.855 --> 00:14:47.690
be able to interpret speech for a wide
00:14:47.690 --> 00:14:48.850
range of people.
00:14:49.140 --> 00:14:51.430
That are not just like talking into it.
00:14:51.430 --> 00:14:53.550
You might talk into your cell phone,
00:14:53.550 --> 00:14:55.425
but they could be shouting from across
00:14:55.425 --> 00:14:56.190
the house.
00:14:56.190 --> 00:14:59.710
Alexa, turn on turn play The Beatles or
00:14:59.710 --> 00:15:00.870
Alexa, what's the temperature?
00:15:01.490 --> 00:15:05.090
And Alexa needs to turn that into text
00:15:05.090 --> 00:15:08.670
and then use other engines in order to
00:15:08.670 --> 00:15:09.880
produce an answer that might be
00:15:09.880 --> 00:15:10.980
relevant question.
00:15:12.790 --> 00:15:13.276
Yeah.
00:15:13.276 --> 00:15:15.530
Chat, ChatGPT is another example, but
00:15:15.530 --> 00:15:17.040
that's quite different, yeah.
00:15:18.220 --> 00:15:23.290
And so for you could Alexa could turn
00:15:23.290 --> 00:15:23.540
that.
00:15:23.540 --> 00:15:25.500
I'm sure that there are ChatGPT apps
00:15:25.500 --> 00:15:26.980
that are connected to Alexa already
00:15:26.980 --> 00:15:29.610
where you can then say, Alexa write my
00:15:29.610 --> 00:15:30.650
essay for me.
00:15:35.280 --> 00:15:36.540
So.
00:15:37.030 --> 00:15:38.660
So for so they had this really
00:15:38.660 --> 00:15:40.030
challenging speech recognition
00:15:40.030 --> 00:15:41.846
algorithm and they realized that the
00:15:41.846 --> 00:15:43.300
state-of-the-art wasn't up to the task.
00:15:43.300 --> 00:15:46.400
So they needed to achieve a certain
00:15:46.400 --> 00:15:48.853
word error rate, which means that the
00:15:48.853 --> 00:15:51.650
rate, the error rate for converting the
00:15:51.650 --> 00:15:54.210
spoken words into text.
00:15:55.090 --> 00:15:57.053
And they were kind of far below that.
00:15:57.053 --> 00:15:59.120
And so first, they started buying
00:15:59.120 --> 00:16:01.590
companies that were developing speech
00:16:01.590 --> 00:16:02.656
recognition engines.
00:16:02.656 --> 00:16:05.280
They started using deep learning.
00:16:05.280 --> 00:16:06.930
They started to try to improve their
00:16:06.930 --> 00:16:10.002
Algorithms, collect data, Annotate the
00:16:10.002 --> 00:16:10.445
data.
00:16:10.445 --> 00:16:13.360
And they had meetings with Jeff Bezos
00:16:13.360 --> 00:16:18.590
and said, we expect that within 10
00:16:18.590 --> 00:16:20.120
years, we're going to achieve the word
00:16:20.120 --> 00:16:21.470
error rate that we need.
00:16:21.470 --> 00:16:23.850
And basically Jeff Bezos, this was like
00:16:23.850 --> 00:16:25.370
his favorite project and he's like,
00:16:25.370 --> 00:16:25.580
well.
00:16:25.630 --> 00:16:27.130
You guys aren't thinking big enough.
00:16:27.130 --> 00:16:29.220
Basically we have tons of money.
00:16:29.220 --> 00:16:31.380
Figure out how you can do it faster.
00:16:31.380 --> 00:16:33.390
And So what they ended up doing was.
00:16:33.470 --> 00:16:37.040
Venting Airbnbs in Boston and other
00:16:37.040 --> 00:16:39.710
places, setting up Alexis, they're
00:16:39.710 --> 00:16:42.488
creating scripts and hiring people to
00:16:42.488 --> 00:16:44.672
walk through these Airbnbs reading the
00:16:44.672 --> 00:16:46.495
scripts, reading a certain script in
00:16:46.495 --> 00:16:49.490
each room so that way they know what is
00:16:49.490 --> 00:16:51.785
said in each room, so they have the
00:16:51.785 --> 00:16:52.355
ground truth.
00:16:52.355 --> 00:16:54.252
The script is basically the ground
00:16:54.252 --> 00:16:56.470
truth, and they can automatically train
00:16:56.470 --> 00:16:59.542
and test their systems on all of these
00:16:59.542 --> 00:17:02.160
people that they on the voices of all
00:17:02.160 --> 00:17:03.446
the people they hired to walk through
00:17:03.446 --> 00:17:04.160
all these different.
00:17:04.220 --> 00:17:04.940
Apartments.
00:17:04.940 --> 00:17:08.370
And so they were able to thousand X
00:17:08.370 --> 00:17:11.100
their training examples and that's how
00:17:11.100 --> 00:17:13.670
they achieved their robust speech
00:17:13.670 --> 00:17:14.820
recognition that they have.
00:17:16.400 --> 00:17:20.160
And so it's kind of the lesson is that
00:17:20.160 --> 00:17:22.430
machine learning is really, it's not
00:17:22.430 --> 00:17:25.062
just about developing algorithms or
00:17:25.062 --> 00:17:28.299
about tuning your losses or tuning your
00:17:28.300 --> 00:17:30.690
parameters, but it's also about data
00:17:30.690 --> 00:17:32.620
preparation and the actual deployment
00:17:32.620 --> 00:17:34.420
of the technology.
00:17:37.840 --> 00:17:39.590
So this.
00:17:40.850 --> 00:17:42.860
So again, like in this class, we're
00:17:42.860 --> 00:17:44.690
going to focus on the Algorithm and
00:17:44.690 --> 00:17:46.610
model development and partly it's
00:17:46.610 --> 00:17:48.370
because there's not time to go out and
00:17:48.370 --> 00:17:49.855
collect lots of data and Annotate it.
00:17:49.855 --> 00:17:51.420
It's extremely time consuming.
00:17:51.420 --> 00:17:53.306
So this is what we can focus on and it
00:17:53.306 --> 00:17:55.360
is the most challenging and technically
00:17:55.360 --> 00:17:58.949
complex part of the development.
00:18:01.360 --> 00:18:03.910
A lot of most, if not all, machine
00:18:03.910 --> 00:18:06.270
learning algorithms fall into this very
00:18:06.270 --> 00:18:07.790
simple diagram.
00:18:08.550 --> 00:18:10.690
Where you have some Raw data, it could
00:18:10.690 --> 00:18:13.330
be text or images or audio or
00:18:13.330 --> 00:18:15.500
temperatures or other information.
00:18:16.520 --> 00:18:19.466
Usually that Raw data is not in a form
00:18:19.466 --> 00:18:22.210
that is very useful for Prediction.
00:18:22.210 --> 00:18:24.502
For example, if you receive an image,
00:18:24.502 --> 00:18:27.010
the intensities of the image are just
00:18:27.010 --> 00:18:28.780
numbers that correspond to how bright
00:18:28.780 --> 00:18:31.410
each pixel is, and individually those
00:18:31.410 --> 00:18:32.860
numbers don't really tell you much at
00:18:32.860 --> 00:18:33.530
all.
00:18:33.530 --> 00:18:35.296
What's more meaningful are the
00:18:35.296 --> 00:18:37.550
patterns, the local patterns of the
00:18:37.550 --> 00:18:37.950
intensities.
00:18:38.810 --> 00:18:40.590
And so you need to have some kind of
00:18:40.590 --> 00:18:44.110
encoding of that Raw data that makes
00:18:44.110 --> 00:18:45.370
the data more meaningful.
00:18:46.340 --> 00:18:49.240
And there's lots of different ways that
00:18:49.240 --> 00:18:50.530
you can create that encoding.
00:18:50.530 --> 00:18:52.315
You can manually define features, you
00:18:52.315 --> 00:18:54.870
could have Trees, you could do
00:18:54.870 --> 00:18:57.060
Probabilistic estimation, or use Deep
00:18:57.060 --> 00:18:57.850
networks.
00:18:58.600 --> 00:18:59.715
Then you have a Decoder.
00:18:59.715 --> 00:19:01.210
The Decoder takes your encoded
00:19:01.210 --> 00:19:03.480
representation and then it tries to
00:19:03.480 --> 00:19:05.060
produce something useful from it.
00:19:05.060 --> 00:19:07.120
Some Prediction that you want classify
00:19:07.120 --> 00:19:10.910
an image or convert the raw audio into
00:19:10.910 --> 00:19:11.390
text.
00:19:12.680 --> 00:19:14.663
Again, there's lots of decoders.
00:19:14.663 --> 00:19:16.790
There's actually not so much complexity
00:19:16.790 --> 00:19:18.910
usually in the decoders, but some of
00:19:18.910 --> 00:19:20.800
the common ones are nearest neighbor or
00:19:20.800 --> 00:19:24.190
Linear decoders, an SVM or logistic,
00:19:24.190 --> 00:19:25.220
logistic regressor.
00:19:26.960 --> 00:19:28.560
Then the decoders will produce some
00:19:28.560 --> 00:19:30.630
Prediction from the encoded Raw data.
00:19:30.630 --> 00:19:32.480
And again, there's lots of different
00:19:32.480 --> 00:19:33.670
kinds of problems out there, so there
00:19:33.670 --> 00:19:34.850
are many different things that you
00:19:34.850 --> 00:19:35.820
might be trying to predict.
00:19:35.820 --> 00:19:37.610
Could be binary labels, you could be
00:19:37.610 --> 00:19:39.710
producing an image, or producing Text,
00:19:39.710 --> 00:19:41.690
or detecting objects.
00:19:42.620 --> 00:19:44.666
When you're training, you'll also have
00:19:44.666 --> 00:19:46.330
some target Labels, so you have
00:19:46.330 --> 00:19:47.000
annotated data.
00:19:47.000 --> 00:19:48.390
If you have a Supervised Algorithm
00:19:48.390 --> 00:19:51.150
anyway, you have some annotated data
00:19:51.150 --> 00:19:52.426
where you have the Raw data and you
00:19:52.426 --> 00:19:53.580
have the labels that you want to
00:19:53.580 --> 00:19:54.390
predict.
00:19:54.390 --> 00:19:56.410
You see how different your target
00:19:56.410 --> 00:19:58.650
Labels are from the predictions, and
00:19:58.650 --> 00:20:00.390
then based on that you.
00:20:01.190 --> 00:20:04.225
Improve your Decoder, and in some cases
00:20:04.225 --> 00:20:06.150
you can then feed that back into your
00:20:06.150 --> 00:20:08.600
Encoder to improve your encoding
00:20:08.600 --> 00:20:10.270
representation.
00:20:10.270 --> 00:20:12.520
So there's a single process that is
00:20:12.520 --> 00:20:14.890
used by almost all of machine learning,
00:20:14.890 --> 00:20:15.990
but there's a lot of different
00:20:15.990 --> 00:20:17.850
variation and a lot of different
00:20:17.850 --> 00:20:18.600
details.
00:20:18.600 --> 00:20:20.414
And that's just because of all the
00:20:20.414 --> 00:20:21.897
different kinds of Raw data that you
00:20:21.897 --> 00:20:23.864
could be dealing with in all the
00:20:23.864 --> 00:20:25.213
different kinds of predictions that you
00:20:25.213 --> 00:20:25.730
could do.
00:20:29.380 --> 00:20:31.157
So that's machine learning.
00:20:31.157 --> 00:20:33.140
That's machine learning in general.
00:20:33.140 --> 00:20:34.470
And so now I'm going to talk a little
00:20:34.470 --> 00:20:36.020
bit about some of the details of the
00:20:36.020 --> 00:20:36.730
course.
00:20:37.780 --> 00:20:40.500
So one of the one of the main course
00:20:40.500 --> 00:20:42.680
objectives is to try to learn how to
00:20:42.680 --> 00:20:44.420
solve problems with machine learning.
00:20:45.110 --> 00:20:47.990
So this involves trying to learn the
00:20:47.990 --> 00:20:50.132
key concepts and methodologies for how
00:20:50.132 --> 00:20:51.140
we can learn from data.
00:20:51.140 --> 00:20:53.170
We're going to learn about a wide
00:20:53.170 --> 00:20:56.230
variety of algorithms and what they're
00:20:56.230 --> 00:20:57.890
strengths and limitations are.
00:20:57.890 --> 00:21:00.110
And we're also going to learn about
00:21:00.110 --> 00:21:02.138
domain specific representations like
00:21:02.138 --> 00:21:04.090
how we can encode images, how we can
00:21:04.090 --> 00:21:06.220
encode text, how we can encode audio.
00:21:06.220 --> 00:21:10.095
And then we want to at the end of this
00:21:10.095 --> 00:21:11.540
have the ability to select the right
00:21:11.540 --> 00:21:13.870
tools for the job so that if you have
00:21:13.870 --> 00:21:15.530
your own custom problem that.
00:21:15.580 --> 00:21:17.180
You want to solve that.
00:21:17.180 --> 00:21:19.270
You're able to make good choices about
00:21:19.270 --> 00:21:22.000
how to represent the data and what
00:21:22.000 --> 00:21:24.160
kinds of algorithms and losses to use,
00:21:24.160 --> 00:21:26.812
how to optimize your algorithms, and
00:21:26.812 --> 00:21:29.140
how to solve your problem.
00:21:31.010 --> 00:21:34.049
So I think this is, I think it's just
00:21:34.050 --> 00:21:36.510
generally interesting to be able to do
00:21:36.510 --> 00:21:38.450
this, but it's also very practical.
00:21:38.450 --> 00:21:41.280
One example is just if you look at it
00:21:41.280 --> 00:21:43.880
in terms of the number of jobs that are
00:21:43.880 --> 00:21:45.930
available in the demand for people that
00:21:45.930 --> 00:21:48.250
have skills in machine learning
00:21:48.250 --> 00:21:50.810
engineering, it's an area that's been
00:21:50.810 --> 00:21:52.260
growing really rapidly.
00:21:52.260 --> 00:21:55.469
So according to this, according to this
00:21:55.470 --> 00:21:55.910
source.
00:21:56.540 --> 00:21:58.220
The.
00:21:58.310 --> 00:22:00.380
Market for global machine learning is
00:22:00.380 --> 00:22:02.860
expected to grow from around 20 billion
00:22:02.860 --> 00:22:07.230
last year to 200 billion in 2029, which
00:22:07.230 --> 00:22:09.080
is a growth rate of about 38%.
00:22:09.080 --> 00:22:11.340
So machine learning engineers are in
00:22:11.340 --> 00:22:13.300
high demand now and they're going to
00:22:13.300 --> 00:22:14.959
likely continue to be in high demand
00:22:14.960 --> 00:22:16.340
over the next several years.
00:22:18.860 --> 00:22:21.540
Another objective of the course is to
00:22:21.540 --> 00:22:23.010
have a better understanding of the real
00:22:23.010 --> 00:22:25.150
life applications and the social
00:22:25.150 --> 00:22:26.758
implications of machine learning.
00:22:26.758 --> 00:22:29.330
So I'll try to talk about actual
00:22:29.330 --> 00:22:30.720
deployments and machine learning and
00:22:30.720 --> 00:22:31.712
several cases.
00:22:31.712 --> 00:22:35.870
We'll talk about some of the ethical
00:22:35.870 --> 00:22:37.479
concerns around machine learning and
00:22:37.480 --> 00:22:40.899
its use, and about the and about
00:22:40.900 --> 00:22:42.390
different application domains.
00:22:42.390 --> 00:22:44.440
So machine learning can do a lot of
00:22:44.440 --> 00:22:45.050
good.
00:22:45.050 --> 00:22:46.800
There's a lot of ways that.
00:22:46.860 --> 00:22:49.170
Already impacts our lives through our
00:22:49.170 --> 00:22:52.330
cameras, through through Smart
00:22:52.330 --> 00:22:54.283
assistants and things like that.
00:22:54.283 --> 00:22:58.193
Of course, technology can also create a
00:22:58.193 --> 00:23:00.453
lot of upheaval and also can do a lot
00:23:00.453 --> 00:23:01.230
of damage.
00:23:01.230 --> 00:23:03.820
And so it's important to understand the
00:23:03.820 --> 00:23:06.390
implications of the technology that we
00:23:06.390 --> 00:23:06.940
develop.
00:23:09.670 --> 00:23:11.630
And then the third is appreciation for
00:23:11.630 --> 00:23:13.160
your own constantly Learning minds.
00:23:13.160 --> 00:23:15.480
So we're humans are still the best
00:23:15.480 --> 00:23:17.180
machine Learners, I would say.
00:23:18.290 --> 00:23:20.490
You're always you're always learning
00:23:20.490 --> 00:23:21.370
new things.
00:23:21.370 --> 00:23:23.170
You're always learning about new
00:23:23.170 --> 00:23:24.140
environments.
00:23:24.140 --> 00:23:27.730
You learn new skills, new facts, new
00:23:27.730 --> 00:23:28.720
concepts.
00:23:28.720 --> 00:23:34.666
And the flexibility of our Learning and
00:23:34.666 --> 00:23:37.140
the breadth of it and the seamlessness
00:23:37.140 --> 00:23:39.040
of it is really amazing.
00:23:39.040 --> 00:23:42.420
And so as you learn how, even though we
00:23:42.420 --> 00:23:46.290
can do pretty cool things with machine
00:23:46.290 --> 00:23:47.940
learning and computer science.
00:23:48.380 --> 00:23:51.140
It's much more so far.
00:23:51.140 --> 00:23:52.880
Machine learning is much more rigid and
00:23:52.880 --> 00:23:55.370
focused and.
00:23:55.520 --> 00:23:56.850
And challenging.
00:23:56.850 --> 00:24:00.675
And so you can think about how did you
00:24:00.675 --> 00:24:02.520
learn to do things, how do you learn to
00:24:02.520 --> 00:24:03.144
make predictions?
00:24:03.144 --> 00:24:05.760
And think about how it may be similar
00:24:05.760 --> 00:24:07.940
or different from the machine learning
00:24:07.940 --> 00:24:09.480
algorithms that we learn about.
00:24:12.410 --> 00:24:13.760
So.
00:24:14.230 --> 00:24:15.800
All right, so now I'm getting more into
00:24:15.800 --> 00:24:18.030
the logistics of the course I
00:24:18.030 --> 00:24:19.280
introduced myself already.
00:24:19.280 --> 00:24:21.100
I want to introduce a few of the tasks
00:24:21.100 --> 00:24:21.660
that are here.
00:24:21.660 --> 00:24:23.380
There's one traveling and one sick.
00:24:24.700 --> 00:24:26.960
But some of them are here, so one is
00:24:26.960 --> 00:24:29.120
Vatsal Chheda.
00:24:29.120 --> 00:24:30.650
Could you introduce yourself?
00:24:32.750 --> 00:24:34.880
And I think I should probably give you
00:24:34.880 --> 00:24:35.570
the mic.
00:24:39.030 --> 00:24:40.740
If I can, sort of.
00:24:46.550 --> 00:24:49.735
Hello everyone my name is Vatsal, I'm
00:24:49.735 --> 00:24:53.290
from India and the like the
00:24:53.290 --> 00:24:54.490
applications of applied machine
00:24:54.490 --> 00:24:55.070
learning are.
00:24:55.070 --> 00:24:56.390
I'm using it in neural machine
00:24:56.390 --> 00:24:58.460
translation from Polish language to
00:24:58.460 --> 00:25:00.630
English language and in various NLP
00:25:00.630 --> 00:25:01.420
projects.
00:25:02.740 --> 00:25:03.540
Thank you.
00:25:05.100 --> 00:25:08.220
I think Josh is out sick today, Weijie.
00:25:14.790 --> 00:25:16.900
Hi everyone, I'm Weijie.
00:25:16.900 --> 00:25:19.235
I'm a second year Masters student and
00:25:19.235 --> 00:25:20.565
I'm from China.
00:25:20.565 --> 00:25:23.920
So I basically do research about
00:25:23.920 --> 00:25:26.860
computer vision and machine learning.
00:25:26.860 --> 00:25:28.510
So nice to meet you all.
00:25:28.510 --> 00:25:29.120
Thank you.
00:25:31.120 --> 00:25:32.440
OK.
00:25:32.440 --> 00:25:34.720
And since each units here, I'll go to
00:25:34.720 --> 00:25:35.570
YouTube next.
00:25:39.540 --> 00:25:41.770
Hi everyone I'm reaching you and I'm a
00:25:41.770 --> 00:25:44.863
first Masters student in the group and
00:25:44.863 --> 00:25:48.420
I came from Shenzhen City located in
00:25:48.420 --> 00:25:51.630
southern China and I'm interested in 3D
00:25:51.630 --> 00:25:53.240
reconstruction and 3D scene
00:25:53.240 --> 00:25:53.850
understanding.
00:25:53.850 --> 00:25:56.160
And during my research I have utilized
00:25:56.160 --> 00:25:59.400
some deep neural network, RESNET and
00:25:59.400 --> 00:26:01.040
multi layer perception and they are
00:26:01.040 --> 00:26:04.010
really useful and can produce pretty
00:26:04.010 --> 00:26:04.910
decent results.
00:26:04.910 --> 00:26:06.550
Looking forward to learning with you.
00:26:06.610 --> 00:26:07.150
Thank you.
00:26:10.220 --> 00:26:12.630
All right and.
00:26:12.990 --> 00:26:16.420
OK, **** is traveling and let's see
00:26:16.420 --> 00:26:17.080
Mington.
00:26:21.990 --> 00:26:24.343
Hi everyone, my name is Min Hongchang
00:26:24.343 --> 00:26:26.820
and I'm a first year masters master
00:26:26.820 --> 00:26:28.790
student in computer science and I'm
00:26:28.790 --> 00:26:33.210
from China and my current research is
00:26:33.210 --> 00:26:35.380
focused on computer vision, especially
00:26:35.380 --> 00:26:38.515
for 3D Vision and generative modeling.
00:26:38.515 --> 00:26:41.690
And I hope you all enjoyed this course.
00:26:41.690 --> 00:26:42.380
Thank you.
00:26:43.010 --> 00:26:45.090
Here and Wentao.
00:26:51.480 --> 00:26:54.630
So on Wentao and I'm a second year
00:26:54.630 --> 00:26:57.080
master student here and last semester
00:26:57.080 --> 00:27:01.095
also tax CS440 if someone take it.
00:27:01.095 --> 00:27:04.510
And my research mainly focus on any
00:27:04.510 --> 00:27:06.380
kind of sequential Prediction like
00:27:06.380 --> 00:27:07.560
natural language understanding,
00:27:07.560 --> 00:27:09.840
national generation, even some
00:27:09.840 --> 00:27:12.670
sequential Prediction in India Vision
00:27:12.670 --> 00:27:15.070
post like trajectory.
00:27:15.130 --> 00:27:17.540
So if you have any question about that
00:27:17.540 --> 00:27:18.600
you can ask me.
00:27:21.050 --> 00:27:22.085
Alright, thank you.
00:27:22.085 --> 00:27:25.510
And I'll have Josh and kitchenette
00:27:25.510 --> 00:27:27.090
introduced themselves.
00:27:28.350 --> 00:27:29.690
When they another time.
00:27:34.360 --> 00:27:34.830
All right.
00:27:34.830 --> 00:27:36.490
So there's.
00:27:37.730 --> 00:27:39.526
So there's three main topics that we're
00:27:39.526 --> 00:27:40.080
going to cover.
00:27:40.080 --> 00:27:43.090
One is Supervised Learning
00:27:43.090 --> 00:27:43.850
Fundamentals.
00:27:43.850 --> 00:27:46.346
So we'll do the next lecture is going
00:27:46.346 --> 00:27:48.890
to be on KNN's and then we'll talk
00:27:48.890 --> 00:27:52.160
about Naive Bayes algorithm and be kind
00:27:52.160 --> 00:27:55.406
of a review of probability as well.
00:27:55.406 --> 00:27:58.170
Then linear and logistic regression
00:27:58.170 --> 00:28:01.180
trees and random forests, and maybe
00:28:01.180 --> 00:28:05.383
boosting ensemble methods, SVMs, neural
00:28:05.383 --> 00:28:07.070
networks and deep neural networks.
00:28:07.900 --> 00:28:09.590
Then the next section, the class, we're
00:28:09.590 --> 00:28:10.880
going to talk about some different
00:28:10.880 --> 00:28:12.010
application domains.
00:28:12.910 --> 00:28:15.447
So we'll talk about computer vision and
00:28:15.447 --> 00:28:19.320
using CNS for computer vision language
00:28:19.320 --> 00:28:21.770
Models, will talk about using the
00:28:21.770 --> 00:28:24.424
transformer networks for vision and
00:28:24.424 --> 00:28:24.930
language.
00:28:24.930 --> 00:28:26.720
This idea that you may have heard about
00:28:26.720 --> 00:28:28.200
called Foundation Models where you
00:28:28.200 --> 00:28:30.789
where you create machine, where you
00:28:30.790 --> 00:28:32.280
train machine learning models on a lot
00:28:32.280 --> 00:28:33.855
of data and then you kind of adapt it
00:28:33.855 --> 00:28:34.565
for other tasks.
00:28:34.565 --> 00:28:37.070
And so we'll talk about task and the
00:28:37.070 --> 00:28:41.800
main adaptation audio as well the
00:28:41.800 --> 00:28:42.500
ethics.
00:28:42.570 --> 00:28:43.120
Issues.
00:28:43.120 --> 00:28:45.600
Really ethical issues relating to
00:28:45.600 --> 00:28:46.310
machine learning.
00:28:47.040 --> 00:28:48.540
And data.
00:28:49.460 --> 00:28:50.620
And then the final section.
00:28:50.620 --> 00:28:53.150
The class is on parent Discovery and
00:28:53.150 --> 00:28:54.990
that focuses more on the unsupervised
00:28:54.990 --> 00:28:56.990
method, so methods for Clustering and
00:28:56.990 --> 00:28:59.290
retrieval, missing data and the
00:28:59.290 --> 00:29:01.290
expectation maximization algorithm
00:29:01.290 --> 00:29:02.530
Density estimation.
00:29:03.590 --> 00:29:06.090
Creating Topic Models, dealing with
00:29:06.090 --> 00:29:08.710
outliers and data visualization.
00:29:08.710 --> 00:29:11.390
CCA, So some of the details might
00:29:11.390 --> 00:29:13.840
change, especially towards the end of
00:29:13.840 --> 00:29:14.360
the class.
00:29:14.360 --> 00:29:16.960
I'm not setting it in stone yet because
00:29:16.960 --> 00:29:20.110
I'll see how things evolve and see if I
00:29:20.110 --> 00:29:22.690
have better ideas later, but at least
00:29:22.690 --> 00:29:26.780
the initial few weeks are more or less
00:29:26.780 --> 00:29:29.090
determined in the schedules online, as
00:29:29.090 --> 00:29:29.800
I'll show.
00:29:32.380 --> 00:29:33.170
So.
00:29:33.260 --> 00:29:33.920
00:29:37.340 --> 00:29:37.600
Don't.
00:29:37.600 --> 00:29:39.410
I don't usually have people cheer when
00:29:39.410 --> 00:29:41.950
I show the show the Assignments.
00:29:43.180 --> 00:29:45.690
Alright, so there's 222 different
00:29:45.690 --> 00:29:46.850
components to your grades.
00:29:46.850 --> 00:29:48.670
There's homeworks and Final projects.
00:29:48.670 --> 00:29:50.700
There's four signed homeworks.
00:29:50.700 --> 00:29:52.640
Those are worth at least 100 points
00:29:52.640 --> 00:29:53.000
each.
00:29:53.710 --> 00:29:56.190
So you can see homework one now that's
00:29:56.190 --> 00:29:59.930
worth 160 points, but the only about
00:29:59.930 --> 00:30:01.413
100 is really expected.
00:30:01.413 --> 00:30:03.830
So for most of the homeworks, there's a
00:30:03.830 --> 00:30:07.762
core of like 100 points that is that is
00:30:07.762 --> 00:30:10.910
more prescripted and then there's like
00:30:10.910 --> 00:30:13.770
additional points for that require more
00:30:13.770 --> 00:30:15.110
independent exploration.
00:30:16.320 --> 00:30:18.550
There's a Final project that's worth
00:30:18.550 --> 00:30:21.940
100 points, and what I'm planning to do
00:30:21.940 --> 00:30:26.270
for that is to select a few.
00:30:26.380 --> 00:30:29.820
A few key challenges, for example,
00:30:29.820 --> 00:30:32.335
something like the Netflix challenge,
00:30:32.335 --> 00:30:34.820
and I'll solicit your input on that.
00:30:34.820 --> 00:30:36.910
And so everyone would have the option
00:30:36.910 --> 00:30:38.540
of doing one of those challenges.
00:30:38.540 --> 00:30:40.660
Or you can just do your own custom
00:30:40.660 --> 00:30:41.180
project.
00:30:43.190 --> 00:30:45.150
So you can kind of.
00:30:47.170 --> 00:30:47.610
There is.
00:30:47.610 --> 00:30:49.055
There's a lot of different students in
00:30:49.055 --> 00:30:50.380
the class from different backgrounds
00:30:50.380 --> 00:30:51.600
and with different interests.
00:30:51.600 --> 00:30:53.870
And so rather than trying to tell
00:30:53.870 --> 00:30:56.200
everybody, make everybody do exactly
00:30:56.200 --> 00:30:57.430
the same thing.
00:30:57.430 --> 00:31:00.129
I generally like to have kind of like
00:31:00.130 --> 00:31:02.240
structure, but options so that if
00:31:02.240 --> 00:31:04.430
you're not interested in something or
00:31:04.430 --> 00:31:07.449
if you're Schedule is really bad for a
00:31:07.450 --> 00:31:09.505
couple weeks, you can.
00:31:09.505 --> 00:31:11.465
There's flexibility built in so that
00:31:11.465 --> 00:31:12.869
you can do the things that make the
00:31:12.870 --> 00:31:15.220
most sense for you so.
00:31:15.380 --> 00:31:16.706
If you're in the three credit version
00:31:16.706 --> 00:31:18.779
of the course, you're graded out of a
00:31:18.780 --> 00:31:21.910
total of 450 points, which means that
00:31:21.910 --> 00:31:24.940
you need to do 4 1/2 of those things.
00:31:24.940 --> 00:31:25.816
But you could.
00:31:25.816 --> 00:31:28.835
You could, for example, do like 3
00:31:28.835 --> 00:31:32.095
homeworks and do all the extra points
00:31:32.095 --> 00:31:33.440
that are available, and then you'd
00:31:33.440 --> 00:31:35.130
probably be done for the semester.
00:31:35.130 --> 00:31:38.386
Or you could do like the bare minimum
00:31:38.386 --> 00:31:43.479
on 3 homeworks, do half and half of
00:31:43.480 --> 00:31:45.530
another homework in the Final project.
00:31:45.590 --> 00:31:47.370
Or any combination, it's up to you.
00:31:48.540 --> 00:31:50.500
And for the four credit version, you
00:31:50.500 --> 00:31:54.682
have to do like everything and a little
00:31:54.682 --> 00:31:55.597
bit more.
00:31:55.597 --> 00:31:59.670
So you have to do the four Assignments
00:31:59.670 --> 00:32:02.967
a little bit of the extra, some of the,
00:32:02.967 --> 00:32:06.730
some of the stretch goals and the Final
00:32:06.730 --> 00:32:07.280
project.
00:32:07.280 --> 00:32:09.020
But again, there's enough opportunity
00:32:09.020 --> 00:32:10.400
that you could just do the four
00:32:10.400 --> 00:32:12.230
homeworks and the stretch goals and be
00:32:12.230 --> 00:32:16.073
done or do the OR do a more basic job.
00:32:16.073 --> 00:32:18.089
There's a little bit of.
00:32:18.150 --> 00:32:20.730
Extra credit available so if you do an
00:32:20.730 --> 00:32:21.510
extra.
00:32:22.400 --> 00:32:24.470
You can earn an additional 15 points
00:32:24.470 --> 00:32:27.080
beyond what's beyond the amount needed
00:32:27.080 --> 00:32:29.150
for 100% of the homework grade.
00:32:29.890 --> 00:32:32.180
So I have in the Syllabus, I have like
00:32:32.180 --> 00:32:33.870
the actual equation for computing your
00:32:33.870 --> 00:32:34.120
grades.
00:32:34.120 --> 00:32:36.310
So if there's any confusion about how
00:32:36.310 --> 00:32:38.127
that works, you can look at that
00:32:38.127 --> 00:32:39.486
equation and it's pretty.
00:32:39.486 --> 00:32:41.130
I think it's pretty straightforward, at
00:32:41.130 --> 00:32:41.970
least once you see that.
00:32:43.280 --> 00:32:47.032
Late policy likewise, I like to be kind
00:32:47.032 --> 00:32:51.537
of flexible and mainly the main reason
00:32:51.537 --> 00:32:53.620
to have late penalties is that I want
00:32:53.620 --> 00:32:56.560
to keep everybody on track and also for
00:32:56.560 --> 00:32:59.950
course logistics have things so that we
00:32:59.950 --> 00:33:01.359
can like kind of grade things and be
00:33:01.360 --> 00:33:02.410
done with them at some point.
00:33:03.770 --> 00:33:06.010
So the late policy is that you get up
00:33:06.010 --> 00:33:07.730
to 10 free late days that you can use
00:33:07.730 --> 00:33:09.030
on any of these Assignments.
00:33:09.790 --> 00:33:12.250
The exception may be the Final project,
00:33:12.250 --> 00:33:13.930
since that's due towards at the end of
00:33:13.930 --> 00:33:15.330
this semester question.
00:33:15.430 --> 00:33:16.840
Cortana Final project.
00:33:16.840 --> 00:33:19.290
Our TV is going to sign us like massive
00:33:19.290 --> 00:33:20.070
kit like.
00:33:27.580 --> 00:33:30.310
So the question is, for the Final
00:33:30.310 --> 00:33:32.600
project, will the TAs create a massive
00:33:32.600 --> 00:33:33.690
GitHub repository?
00:33:34.400 --> 00:33:36.380
And share it?
00:33:36.380 --> 00:33:37.710
Or do you create your own?
00:33:38.940 --> 00:33:39.730
So.
00:33:40.770 --> 00:33:43.070
Most likely the TAs will not create
00:33:43.070 --> 00:33:44.950
anything for the Final project.
00:33:44.950 --> 00:33:47.520
So you would because there many people
00:33:47.520 --> 00:33:49.920
may do their custom projects and then
00:33:49.920 --> 00:33:52.386
you're doing you're working with it
00:33:52.386 --> 00:33:54.010
could be online, you could start with
00:33:54.010 --> 00:33:56.260
an online repository or do it from
00:33:56.260 --> 00:33:58.150
scratch and likewise for the
00:33:58.150 --> 00:33:59.040
challenges.
00:33:59.120 --> 00:34:02.660
And if we pick Kaggle challenges for
00:34:02.660 --> 00:34:04.440
example, there's some, there's a little
00:34:04.440 --> 00:34:06.000
bit of like infrastructure around
00:34:06.000 --> 00:34:08.240
those, but mostly for the final
00:34:08.240 --> 00:34:10.050
projects, I want it to be more
00:34:10.050 --> 00:34:10.800
open-ended.
00:34:10.800 --> 00:34:12.100
So yeah.
00:34:13.960 --> 00:34:15.690
And how did this?
00:34:15.690 --> 00:34:17.070
Yeah.
00:34:19.100 --> 00:34:21.351
So the Final project can be done in a
00:34:21.351 --> 00:34:23.037
group or it can be if you so.
00:34:23.037 --> 00:34:25.310
If you do a custom project it has I
00:34:25.310 --> 00:34:27.145
want it has to be in a group.
00:34:27.145 --> 00:34:29.060
If you do one of the challenges you can
00:34:29.060 --> 00:34:30.482
either do it individually or in a
00:34:30.482 --> 00:34:30.640
group.
00:34:32.250 --> 00:34:33.380
Up to four.
00:34:34.740 --> 00:34:35.380
Yeah.
00:34:37.680 --> 00:34:40.490
OK, so back to the late policy.
00:34:40.490 --> 00:34:41.690
So the.
00:34:42.110 --> 00:34:43.835
So you get up to 10 free late days.
00:34:43.835 --> 00:34:45.460
You can use them on any combination of
00:34:45.460 --> 00:34:45.678
projects.
00:34:45.678 --> 00:34:47.400
You can be 10 days late for one
00:34:47.400 --> 00:34:49.640
project, you can be 5 days late for two
00:34:49.640 --> 00:34:50.830
projects, and so on.
00:34:51.810 --> 00:34:54.860
And then there's a 5 point penalty per
00:34:54.860 --> 00:34:55.900
day after that.
00:34:56.860 --> 00:34:58.510
So for example, if you would have
00:34:58.510 --> 00:35:02.422
earned 110 points, but your five days
00:35:02.422 --> 00:35:03.900
late and you only had three late days
00:35:03.900 --> 00:35:05.580
left, then you would earn 100 points
00:35:05.580 --> 00:35:06.090
instead.
00:35:07.170 --> 00:35:08.920
And then the project.
00:35:08.920 --> 00:35:11.945
Every assignment, though, has to every
00:35:11.945 --> 00:35:13.770
homework has to be submitted within two
00:35:13.770 --> 00:35:15.655
weeks of the due date to receive any
00:35:15.655 --> 00:35:15.930
points.
00:35:15.930 --> 00:35:19.114
So you can't submit homework one like
00:35:19.114 --> 00:35:20.786
three weeks late and still get points
00:35:20.786 --> 00:35:21.350
for it.
00:35:21.350 --> 00:35:23.201
And part of the reason for that is that
00:35:23.201 --> 00:35:25.769
I don't want, I don't want any students
00:35:25.769 --> 00:35:28.330
to just get like chronically behind
00:35:28.330 --> 00:35:29.583
where you're submitting every
00:35:29.583 --> 00:35:31.040
assignment two or three weeks late,
00:35:31.040 --> 00:35:32.050
because I've seen that happen.
00:35:32.050 --> 00:35:34.570
And then I've had students build up
00:35:34.570 --> 00:35:36.480
like 60 late days by the end of this
00:35:36.480 --> 00:35:37.280
semester.
00:35:37.930 --> 00:35:40.340
So it's better just to move on if
00:35:40.340 --> 00:35:41.330
you're 2 weeks late.
00:35:43.450 --> 00:35:44.460
00:35:46.510 --> 00:35:46.900
Homework.
00:35:48.990 --> 00:35:51.270
The Final project I have not yet
00:35:51.270 --> 00:35:53.690
decided, but you definitely can't be
00:35:53.690 --> 00:35:54.875
the Final project.
00:35:54.875 --> 00:35:57.890
I might allow one or two late days, but
00:35:57.890 --> 00:35:59.880
probably not more than that because
00:35:59.880 --> 00:36:02.340
it's at the it's due on the last day of
00:36:02.340 --> 00:36:04.030
the semester, so there's not.
00:36:04.030 --> 00:36:05.920
So there needs to be time for grading
00:36:05.920 --> 00:36:07.940
and also time for you to do your exams.
00:36:10.780 --> 00:36:13.350
Alright, so Covid's Masks and sickness,
00:36:13.350 --> 00:36:15.270
so I do like it.
00:36:15.270 --> 00:36:16.880
If you come to I think it's a good idea
00:36:16.880 --> 00:36:18.290
to come to lecture whenever you can.
00:36:18.290 --> 00:36:19.980
I realize it's early in the morning for
00:36:19.980 --> 00:36:22.770
many people, but it's probably the best
00:36:22.770 --> 00:36:26.865
way to keep you on Schedule and even if
00:36:26.865 --> 00:36:28.870
you just come and.
00:36:29.220 --> 00:36:32.179
And our, I don't know, doing doing
00:36:32.180 --> 00:36:32.881
other work or whatever.
00:36:32.881 --> 00:36:34.350
At least you're at least you're like
00:36:34.350 --> 00:36:35.750
keeping tabs on what's going on in the
00:36:35.750 --> 00:36:36.410
course.
00:36:36.970 --> 00:36:41.190
But everything will be recorded, so
00:36:41.190 --> 00:36:44.090
it's you're also perfectly capable of
00:36:44.090 --> 00:36:46.040
catching up on anything that you missed
00:36:46.040 --> 00:36:46.547
online.
00:36:46.547 --> 00:36:49.720
So if you're well, do come to lectures
00:36:49.720 --> 00:36:51.820
and in office hours if you have
00:36:51.820 --> 00:36:52.690
something to discuss.
00:36:53.610 --> 00:36:55.140
Master optional.
00:36:55.260 --> 00:36:58.440
If you're sick, if you're sick, do you
00:36:58.440 --> 00:37:00.120
stay home because nobody else wants to
00:37:00.120 --> 00:37:00.570
get sick?
00:37:01.560 --> 00:37:03.860
You never need to show any proof of
00:37:03.860 --> 00:37:06.910
illness or you never need to ask me for
00:37:06.910 --> 00:37:08.160
permission to miss a class.
00:37:08.160 --> 00:37:10.880
You can just miss it and then you can
00:37:10.880 --> 00:37:12.490
watch the recording.
00:37:12.490 --> 00:37:16.163
So the lectures will be recorded so you
00:37:16.163 --> 00:37:17.480
can always catch up on them.
00:37:17.480 --> 00:37:19.970
As I was saying also the exams are
00:37:19.970 --> 00:37:22.039
planned to be on Prairie learn, so you
00:37:22.040 --> 00:37:23.496
would be able to take them at home.
00:37:23.496 --> 00:37:24.874
In fact, you won't be able to take them
00:37:24.874 --> 00:37:27.080
in the lecture, you will need to take
00:37:27.080 --> 00:37:29.970
them at home and the exams will be open
00:37:29.970 --> 00:37:30.270
book.
00:37:32.600 --> 00:37:34.910
Doesn't necessarily mean they're easy.
00:37:36.460 --> 00:37:36.810
Open.
00:37:44.860 --> 00:37:46.420
No, it's not.
00:37:46.420 --> 00:37:48.810
All right, so.
00:37:48.880 --> 00:37:50.960
For their homeworks, you will implement
00:37:50.960 --> 00:37:52.870
and apply machine learning methods in
00:37:52.870 --> 00:37:53.810
Jupiter notebooks.
00:37:55.020 --> 00:37:55.895
I'll show you.
00:37:55.895 --> 00:37:57.640
I'll go through the homework on the in
00:37:57.640 --> 00:38:00.450
the next lecture on Thursday, but
00:38:00.450 --> 00:38:02.510
you'll see that the basic structure is
00:38:02.510 --> 00:38:04.630
that you have a.
00:38:05.260 --> 00:38:07.165
You have like a main assignment page
00:38:07.165 --> 00:38:10.220
and then there's starter code which is
00:38:10.220 --> 00:38:12.390
pretty minimal but just enough to give
00:38:12.390 --> 00:38:13.840
some structure and to load the data
00:38:13.840 --> 00:38:14.460
that you need.
00:38:15.220 --> 00:38:18.130
And then there's a.
00:38:18.760 --> 00:38:20.775
So they're like places where you would
00:38:20.775 --> 00:38:22.280
where you would write the code for each
00:38:22.280 --> 00:38:24.980
section, and then there there's a
00:38:24.980 --> 00:38:27.970
report template which is template for
00:38:27.970 --> 00:38:30.340
reporting the results of your
00:38:30.340 --> 00:38:31.040
Algorithms.
00:38:32.160 --> 00:38:34.300
And then there's also like a tips page
00:38:34.300 --> 00:38:36.660
which has common code and some other
00:38:36.660 --> 00:38:38.110
suggestions about the assignment.
00:38:41.270 --> 00:38:43.490
Alright, so this is the main Learning
00:38:43.490 --> 00:38:45.490
resource, the website, I mean this is
00:38:45.490 --> 00:38:47.170
kind of the central repository of
00:38:47.170 --> 00:38:48.230
everything that you need.
00:38:51.110 --> 00:38:53.190
So that showed up there.
00:38:53.190 --> 00:38:57.530
OK, so you can find it from my home
00:38:57.530 --> 00:38:58.030
page.
00:38:58.030 --> 00:38:59.427
And also it's the.
00:38:59.427 --> 00:39:01.405
It's just like the standard location.
00:39:01.405 --> 00:39:04.076
So if you do like
00:39:04.076 --> 00:39:06.909
courses.endure.illinois.edu/CS 441.
00:39:07.630 --> 00:39:09.320
I think even that will go.
00:39:09.320 --> 00:39:11.420
Yeah, even that will just go to spring
00:39:11.420 --> 00:39:12.090
2023.
00:39:12.880 --> 00:39:15.895
So you can see there's a Syllabus.
00:39:15.895 --> 00:39:18.230
There's obviously no lecture recordings
00:39:18.230 --> 00:39:20.740
yet, but once they should show up at
00:39:20.740 --> 00:39:23.320
this location on media space.
00:39:24.770 --> 00:39:26.700
There's the CampusWire.
00:39:26.700 --> 00:39:28.160
You can sign up with this code.
00:39:28.160 --> 00:39:30.640
I'll probably enroll everybody at some
00:39:30.640 --> 00:39:32.575
point this week that's registered for
00:39:32.575 --> 00:39:33.190
the class.
00:39:34.350 --> 00:39:38.615
The submission for Canvas, the
00:39:38.615 --> 00:39:39.250
Textbook.
00:39:39.250 --> 00:39:42.400
So I don't really teach out of a
00:39:42.400 --> 00:39:46.160
textbook, but this book is quite good.
00:39:46.160 --> 00:39:47.620
Applied machine Learning by David
00:39:47.620 --> 00:39:49.350
Forsyth, the professor here.
00:39:49.350 --> 00:39:50.916
He actually created the first version
00:39:50.916 --> 00:39:54.685
of this course and I do look at the
00:39:54.685 --> 00:39:55.960
Textbook to make sure I'm not
00:39:55.960 --> 00:39:57.460
forgetting anything really important.
00:39:58.420 --> 00:40:01.650
And it's useful to have like another
00:40:01.650 --> 00:40:03.340
perspective on some of the topics I'm
00:40:03.340 --> 00:40:06.860
teaching, or to have like more like
00:40:06.860 --> 00:40:07.850
background reading.
00:40:08.930 --> 00:40:12.560
So I'll this is the schedule of topics,
00:40:12.560 --> 00:40:15.225
so you can see that right now we're in
00:40:15.225 --> 00:40:16.300
the introduction.
00:40:16.300 --> 00:40:18.750
I've got the PowerPoint slides and the
00:40:18.750 --> 00:40:19.375
PDF here.
00:40:19.375 --> 00:40:22.490
I generally try to put them online
00:40:22.490 --> 00:40:24.750
before I teach, but I can't guarantee
00:40:24.750 --> 00:40:25.250
it.
00:40:27.280 --> 00:40:29.940
There is a I've got a bunch of
00:40:29.940 --> 00:40:31.420
tutorials here.
00:40:31.420 --> 00:40:33.310
These were originally created for the
00:40:33.310 --> 00:40:35.050
computational photography course, but
00:40:35.050 --> 00:40:37.680
they're pretty general, so one is on
00:40:37.680 --> 00:40:40.630
using Jupiter notebooks, one is on
00:40:40.630 --> 00:40:42.590
Numpy, and another is on linear
00:40:42.590 --> 00:40:46.150
algebra, and they're really good, so I
00:40:46.150 --> 00:40:48.420
would recommend watching those they
00:40:48.420 --> 00:40:48.710
have.
00:40:48.710 --> 00:40:51.860
It's a video, but some of them, like
00:40:51.860 --> 00:40:53.770
this one for Jupiter notebook, also
00:40:53.770 --> 00:40:55.780
comes with a notebook and they'll like
00:40:55.780 --> 00:40:57.380
pause and give you time to try things
00:40:57.380 --> 00:40:58.050
out yourself.
00:40:59.110 --> 00:41:00.540
So those are worth reviewing,
00:41:00.540 --> 00:41:02.570
especially if you're not familiar with
00:41:02.570 --> 00:41:04.230
Jupiter notebooks, if you're not
00:41:04.230 --> 00:41:08.290
familiar with Python, or if you've feel
00:41:08.290 --> 00:41:09.620
like you need a refresher on linear
00:41:09.620 --> 00:41:10.130
algebra.
00:41:11.900 --> 00:41:13.900
The Assignments are linked to here, so
00:41:13.900 --> 00:41:15.960
if you click on that you can it should
00:41:15.960 --> 00:41:17.170
take you to homework one.
00:41:17.890 --> 00:41:21.480
And you can check that out if you're
00:41:21.480 --> 00:41:22.770
interested.
00:41:22.770 --> 00:41:24.770
Like I said, I'm going to review that
00:41:24.770 --> 00:41:26.082
on Thursday.
00:41:26.082 --> 00:41:30.445
And that first homework mainly is the
00:41:30.445 --> 00:41:32.450
that covers the topics that are taught
00:41:32.450 --> 00:41:33.650
in the first three lectures.
00:41:33.650 --> 00:41:34.830
KNN maybes.
00:41:35.460 --> 00:41:37.860
And Linear Logistic Regression.
00:41:39.990 --> 00:41:41.250
So.
00:41:42.260 --> 00:41:42.630
Yes.
00:41:42.630 --> 00:41:43.770
So the other thing.
00:41:45.510 --> 00:41:46.440
Show you the Syllabus.
00:41:46.440 --> 00:41:47.520
So here's the Syllabus.
00:41:48.430 --> 00:41:49.755
Do you read it on your own?
00:41:49.755 --> 00:41:52.660
You can see more or less the same
00:41:52.660 --> 00:41:55.240
information that I just gave you, but
00:41:55.240 --> 00:41:57.460
it has the, I guess a little bit more
00:41:57.460 --> 00:41:58.322
detail on some things.
00:41:58.322 --> 00:41:59.950
It has the greeting equations.
00:42:00.880 --> 00:42:03.680
So the also has grade Thresholds, so.
00:42:04.570 --> 00:42:06.001
If you reach these graded Thresholds
00:42:06.001 --> 00:42:07.507
then you're guaranteed that grade.
00:42:07.507 --> 00:42:10.219
So if you get like a 97 you will
00:42:10.220 --> 00:42:11.480
definitely get an A plus.
00:42:11.480 --> 00:42:13.940
I generally will look at the grade
00:42:13.940 --> 00:42:16.560
distribution at the end of the semester
00:42:16.560 --> 00:42:19.790
and if it seems warranted then I might
00:42:19.790 --> 00:42:21.710
lower the Thresholds but I won't raise
00:42:21.710 --> 00:42:22.000
them.
00:42:22.000 --> 00:42:24.147
So if you get like a 90, you're
00:42:24.147 --> 00:42:24.953
guaranteed a minus.
00:42:24.953 --> 00:42:27.209
It might be that if you get an 89 you
00:42:27.210 --> 00:42:30.444
could also get an A minus, but I it
00:42:30.444 --> 00:42:31.205
depends on.
00:42:31.205 --> 00:42:34.110
It depends on like I have to see.
00:42:34.160 --> 00:42:36.130
Have to at the end of the semester so I
00:42:36.130 --> 00:42:38.080
don't curve any individual Assignments
00:42:38.080 --> 00:42:38.500
I do.
00:42:38.500 --> 00:42:39.820
I can curve.
00:42:40.770 --> 00:42:43.670
The final grades, but usually it will
00:42:43.670 --> 00:42:45.480
not change very much.
00:42:45.480 --> 00:42:47.780
So just think of these as your targets.
00:42:49.530 --> 00:42:51.320
Late policy explained.
00:42:52.810 --> 00:42:55.650
I think all of this, yeah.
00:42:55.650 --> 00:42:59.340
So I guess it's worth noting that this
00:42:59.340 --> 00:43:01.420
is the first time I've taught this
00:43:01.420 --> 00:43:04.740
course, and so I.
00:43:05.490 --> 00:43:08.660
I'm I wanna keep some flexibility
00:43:08.660 --> 00:43:09.255
options open.
00:43:09.255 --> 00:43:11.060
I'm going to be soliciting feedback at
00:43:11.060 --> 00:43:11.850
various points.
00:43:12.670 --> 00:43:16.480
I may course correct, so I'll do so in
00:43:16.480 --> 00:43:18.530
a way that's not disruptive and give
00:43:18.530 --> 00:43:20.320
you as much notice as possible.
00:43:20.320 --> 00:43:22.050
But I don't want to set.
00:43:22.050 --> 00:43:23.410
I don't want to.
00:43:24.670 --> 00:43:26.680
To be completely inflexible just so
00:43:26.680 --> 00:43:28.820
that I can make improvements that may
00:43:28.820 --> 00:43:30.110
benefit your experience.
00:43:32.170 --> 00:43:35.470
Alright, so yeah, that's enough of
00:43:35.470 --> 00:43:36.182
reviewing this.
00:43:36.182 --> 00:43:38.120
So you should read all of this stuff,
00:43:38.120 --> 00:43:40.620
but I don't need to do it right now.
00:43:42.800 --> 00:43:44.400
Alright, I think I talked about that
00:43:44.400 --> 00:43:45.310
stuff.
00:43:45.310 --> 00:43:50.900
OK, so the office hours, I do have them
00:43:50.900 --> 00:43:53.270
ready, but I'll create a post on
00:43:53.270 --> 00:43:55.120
CampusWire that will tell you what the
00:43:55.120 --> 00:43:56.580
office hours and pin them.
00:43:56.580 --> 00:43:58.340
The office hours will start next week.
00:43:59.390 --> 00:44:01.380
And then, like I said, the reading.
00:44:01.380 --> 00:44:03.010
The main Readings are the applied
00:44:03.010 --> 00:44:05.260
machine learning book by David Forsyth.
00:44:09.660 --> 00:44:12.370
Alright, so Academic Integrity.
00:44:12.470 --> 00:44:13.190
00:44:14.040 --> 00:44:15.860
It's OK to discuss your homework with
00:44:15.860 --> 00:44:16.750
classmates.
00:44:16.750 --> 00:44:20.780
And actually, if you and somebody else
00:44:20.780 --> 00:44:23.100
finish your homework, I would actually
00:44:23.100 --> 00:44:26.585
encourage you to like to like, discuss
00:44:26.585 --> 00:44:28.304
like what were your results?
00:44:28.304 --> 00:44:30.520
And if your results don't match, then
00:44:30.520 --> 00:44:31.880
find out what they did.
00:44:31.880 --> 00:44:34.570
I'm OK with that, but if you haven't
00:44:34.570 --> 00:44:36.330
done it yet, don't like look at their
00:44:36.330 --> 00:44:39.130
code so that you can do the exact same
00:44:39.130 --> 00:44:40.316
thing that they did.
00:44:40.316 --> 00:44:41.990
So I think, like, there's actually a
00:44:41.990 --> 00:44:43.850
lot of educational value in learning
00:44:43.850 --> 00:44:45.180
from other students and.
00:44:45.890 --> 00:44:47.392
Sometimes there's going to be more than
00:44:47.392 --> 00:44:49.470
one way to implement something, and one
00:44:49.470 --> 00:44:50.840
way will lead to better results than
00:44:50.840 --> 00:44:51.300
another.
00:44:51.950 --> 00:44:55.280
And it's worth finding out what that
00:44:55.280 --> 00:44:56.537
better way is.
00:44:56.537 --> 00:44:59.980
But you should mainly do things
00:44:59.980 --> 00:45:04.070
independently for the Assignments.
00:45:04.070 --> 00:45:05.710
Like I said, for the Final project you
00:45:05.710 --> 00:45:07.300
can work in groups, but for the regular
00:45:07.300 --> 00:45:08.570
Assignments it should be mostly
00:45:08.570 --> 00:45:09.310
independent work.
00:45:09.950 --> 00:45:12.000
And any consultation with others is
00:45:12.000 --> 00:45:14.220
should be aimed at further enhancing
00:45:14.220 --> 00:45:16.700
your what you learn rather than
00:45:16.700 --> 00:45:17.800
bypassing it.
00:45:19.730 --> 00:45:21.322
You can look at Stack Overflow.
00:45:21.322 --> 00:45:23.036
You can get ideas from online.
00:45:23.036 --> 00:45:25.700
At the end of the template there's a
00:45:25.700 --> 00:45:27.940
section where you say what other
00:45:27.940 --> 00:45:29.340
sources you have.
00:45:29.340 --> 00:45:32.031
And so if you talk to anybody about the
00:45:32.031 --> 00:45:33.234
course, I mean not the course.
00:45:33.234 --> 00:45:34.439
If you talk to anybody about the
00:45:34.440 --> 00:45:36.132
assignment or looked at Stack Overflow
00:45:36.132 --> 00:45:38.410
or whatever, you should just list those
00:45:38.410 --> 00:45:38.960
things there.
00:45:38.960 --> 00:45:39.280
So.
00:45:40.070 --> 00:45:42.600
As you're doing the assignment if you.
00:45:42.810 --> 00:45:44.140
If you have like.
00:45:45.780 --> 00:45:48.190
If you end up using other resources,
00:45:48.190 --> 00:45:49.840
just write them down South that you
00:45:49.840 --> 00:45:50.870
remember to put them there.
00:45:52.500 --> 00:45:55.470
So you shouldn't like copy any code
00:45:55.470 --> 00:45:56.570
from anybody.
00:45:56.570 --> 00:45:58.530
So you should never be like claiming
00:45:58.530 --> 00:46:00.250
credit for something that somebody else
00:46:00.250 --> 00:46:01.510
wrote, even if you typed it out
00:46:01.510 --> 00:46:02.880
yourself.
00:46:02.880 --> 00:46:05.050
And you should not use external
00:46:05.050 --> 00:46:06.620
resources but without acknowledging
00:46:06.620 --> 00:46:06.840
them.
00:46:07.640 --> 00:46:08.950
So if you're not sure if something's
00:46:08.950 --> 00:46:11.900
OK, you can just ask and as long as you
00:46:11.900 --> 00:46:14.279
acknowledge all your sources then you
00:46:14.279 --> 00:46:16.230
can then you're safe.
00:46:16.230 --> 00:46:16.830
So.
00:46:17.610 --> 00:46:19.865
Even if even if what you did is
00:46:19.865 --> 00:46:22.685
something that I would consider to be
00:46:22.685 --> 00:46:25.524
like 2 too much copying or something,
00:46:25.524 --> 00:46:27.290
if you just if you disclose it and
00:46:27.290 --> 00:46:28.410
you're acknowledgements, it's not
00:46:28.410 --> 00:46:32.031
cheating, it's just not would not be
00:46:32.031 --> 00:46:33.840
like doing enough I guess.
00:46:36.840 --> 00:46:38.110
So.
00:46:39.080 --> 00:46:39.840
Alright.
00:46:39.840 --> 00:46:42.800
So yeah, in terms of Prerequisites, so
00:46:42.800 --> 00:46:44.380
I'm going to assume that you've had
00:46:44.380 --> 00:46:46.140
some kind of probability and statistics
00:46:46.140 --> 00:46:46.990
before.
00:46:46.990 --> 00:46:49.240
I will review it a little bit when I
00:46:49.240 --> 00:46:52.790
talk about Naive Bayes, but it's really
00:46:52.790 --> 00:46:56.060
like a 20 or 30 minute review to what
00:46:56.060 --> 00:46:57.990
would normally be a course.
00:46:57.990 --> 00:47:01.019
So it's not meant to replace replace
00:47:01.020 --> 00:47:03.130
having taken probably probability or
00:47:03.130 --> 00:47:04.905
statistics and that's not supposed to
00:47:04.905 --> 00:47:07.890
be stages, it's stats I think.
00:47:08.460 --> 00:47:11.610
And the linear algebra.
00:47:11.610 --> 00:47:13.410
So again, I'm going to assume that you
00:47:13.410 --> 00:47:15.830
things like matrix multiplication, what
00:47:15.830 --> 00:47:18.070
inverses, stuff like that.
00:47:19.280 --> 00:47:23.030
If not, if not, then you should
00:47:23.030 --> 00:47:26.010
probably learn it first, but you can
00:47:26.010 --> 00:47:27.590
review it using the tutorial.
00:47:28.450 --> 00:47:30.930
And then also assume some knowledge of
00:47:30.930 --> 00:47:33.230
calculus, like if I take a derivative,
00:47:33.230 --> 00:47:33.620
what is?
00:47:33.620 --> 00:47:35.280
And you kind of know you how to take
00:47:35.280 --> 00:47:36.480
derivatives yourself.
00:47:37.530 --> 00:47:40.490
And if you have experience with Python,
00:47:40.490 --> 00:47:41.990
that will help, because we'll be doing
00:47:41.990 --> 00:47:44.440
the assignments in Python, but it's not
00:47:44.440 --> 00:47:46.920
necessary, I mean I myself.
00:47:48.780 --> 00:47:49.820
Don't.
00:47:49.820 --> 00:47:52.490
I've used Python a bit.
00:47:52.560 --> 00:47:55.750
And I find it pretty easy to do the
00:47:55.750 --> 00:47:58.590
coding, but I'm not like an expert in
00:47:58.590 --> 00:48:00.404
Python by any means, so it's not.
00:48:00.404 --> 00:48:02.016
It's not extremely challenging coding
00:48:02.016 --> 00:48:03.257
in my opinion.
00:48:03.257 --> 00:48:06.823
So if you're generally capable of
00:48:06.823 --> 00:48:08.170
coding, then you're capable of picking
00:48:08.170 --> 00:48:09.890
up Python And doing it, but if you
00:48:09.890 --> 00:48:10.810
already know, it will help.
00:48:12.290 --> 00:48:14.090
And then like I said, if watch the
00:48:14.090 --> 00:48:16.820
tutorials if you would like to review
00:48:16.820 --> 00:48:17.740
those concepts.
00:48:20.990 --> 00:48:24.290
So I also want to just briefly mention
00:48:24.290 --> 00:48:25.840
how this course is different from some
00:48:25.840 --> 00:48:27.280
others, because that's a really common
00:48:27.280 --> 00:48:28.465
question.
00:48:28.465 --> 00:48:32.320
So one of the other main courses is
00:48:32.320 --> 00:48:34.610
446, which is just called machine
00:48:34.610 --> 00:48:35.120
learning.
00:48:36.080 --> 00:48:38.490
So this course, when I say this course
00:48:38.490 --> 00:48:40.180
here, I mean this one that we're in
00:48:40.180 --> 00:48:43.078
right now, 441 this course provides a
00:48:43.078 --> 00:48:45.160
foundation for ML practice, while I
00:48:45.160 --> 00:48:47.370
would say 446 provides more of a
00:48:47.370 --> 00:48:48.700
foundation for machine learning
00:48:48.700 --> 00:48:49.630
research.
00:48:49.630 --> 00:48:52.710
And so compared to 446, we will have
00:48:52.710 --> 00:48:55.660
less theory, fewer derivations, less
00:48:55.660 --> 00:48:59.440
focus on optimization and more focus on
00:48:59.440 --> 00:49:01.610
how you represent things on data
00:49:01.610 --> 00:49:03.300
representations and developing
00:49:03.300 --> 00:49:05.510
applications and.
00:49:05.570 --> 00:49:07.140
Application examples.
00:49:07.140 --> 00:49:08.620
So it doesn't mean that we don't have
00:49:08.620 --> 00:49:10.630
any theory, but it's all everything in
00:49:10.630 --> 00:49:12.790
the course is geared towards making you
00:49:12.790 --> 00:49:15.860
a good machine learning practitioner
00:49:15.860 --> 00:49:19.035
rather than making you a good machine
00:49:19.035 --> 00:49:20.340
learning researcher.
00:49:20.340 --> 00:49:22.260
There are just different focuses.
00:49:23.090 --> 00:49:24.580
Also, if you look at the syllabus for
00:49:24.580 --> 00:49:27.090
446, you'll see that there are some
00:49:27.090 --> 00:49:28.770
topics that are in common and then
00:49:28.770 --> 00:49:30.540
there's others that are different so.
00:49:31.660 --> 00:49:33.740
But you can just, like look at doing an
00:49:33.740 --> 00:49:36.490
AB comparison on the syllabi once if
00:49:36.490 --> 00:49:38.390
the 446 Syllabus is available.
00:49:38.390 --> 00:49:40.700
There's also an online version of this
00:49:40.700 --> 00:49:42.640
course that's currently taught by Marco
00:49:42.640 --> 00:49:43.160
Morales.
00:49:45.060 --> 00:49:48.220
This is a complete redesign for this
00:49:48.220 --> 00:49:48.610
semester.
00:49:48.610 --> 00:49:52.500
My version is a total redesign, so they
00:49:52.500 --> 00:49:54.680
cover similar topics but in different
00:49:54.680 --> 00:49:55.170
ways.
00:49:57.340 --> 00:50:00.820
Assignment wise, this course has fewer
00:50:00.820 --> 00:50:02.700
and larger homeworks that are a little
00:50:02.700 --> 00:50:04.550
bit more open-ended.
00:50:06.450 --> 00:50:08.780
And also as a Final project and exams,
00:50:08.780 --> 00:50:10.820
while the online version, at least the
00:50:10.820 --> 00:50:13.780
last one I saw has like 11 homeworks
00:50:13.780 --> 00:50:16.210
that are a little bit more scripted and
00:50:16.210 --> 00:50:20.100
it has quizzes and it's just a.
00:50:21.320 --> 00:50:23.050
Very different in the kind of
00:50:23.050 --> 00:50:23.500
structure.
00:50:25.420 --> 00:50:26.090
And.
00:50:27.300 --> 00:50:29.720
And I would say that compared to the
00:50:29.720 --> 00:50:31.660
online one, the one that I'm teaching
00:50:31.660 --> 00:50:33.870
now focuses more on a conceptual
00:50:33.870 --> 00:50:37.300
understanding of the techniques and on
00:50:37.300 --> 00:50:38.850
how machine learning is used today.
00:50:40.730 --> 00:50:43.440
And then one more example is there's a
00:50:43.440 --> 00:50:45.005
deep Learning course for computer
00:50:45.005 --> 00:50:45.710
vision.
00:50:45.710 --> 00:50:48.566
So that's focused on, as it says, deep
00:50:48.566 --> 00:50:50.400
learning and computer vision, where
00:50:50.400 --> 00:50:52.916
this course is not only focused on deep
00:50:52.916 --> 00:50:54.850
learning, I teach a variety of
00:50:54.850 --> 00:50:56.955
different techniques, including deep
00:50:56.955 --> 00:50:58.090
learning to some extent.
00:50:58.770 --> 00:51:01.850
And also talk about a broader variety
00:51:01.850 --> 00:51:02.890
of Application Domains.
00:51:04.570 --> 00:51:06.482
So you may be wondering whether you
00:51:06.482 --> 00:51:07.310
take the course.
00:51:07.310 --> 00:51:09.029
I would say you should take the course
00:51:09.030 --> 00:51:10.150
if you want to learn how to apply
00:51:10.150 --> 00:51:10.760
machine learning.
00:51:11.810 --> 00:51:13.626
If you like coding based homeworks and
00:51:13.626 --> 00:51:15.280
you're also OK with math, then you
00:51:15.280 --> 00:51:16.863
might want to take the course.
00:51:16.863 --> 00:51:20.210
Also, you need to be willing to spend a
00:51:20.210 --> 00:51:21.930
significant amount of time, so I would
00:51:21.930 --> 00:51:23.700
say it would probably take 10 to 12
00:51:23.700 --> 00:51:24.760
hours per week.
00:51:24.760 --> 00:51:28.103
If you don't know if you're catching up
00:51:28.103 --> 00:51:29.610
on a lot of things then it could take
00:51:29.610 --> 00:51:31.806
even longer or depending on how fast of
00:51:31.806 --> 00:51:33.290
a programmer you are and things like
00:51:33.290 --> 00:51:33.840
that.
00:51:33.840 --> 00:51:37.470
But this is my estimate, so it is a
00:51:37.470 --> 00:51:38.450
it's not a light course.
00:51:40.230 --> 00:51:42.640
Don't take the course if so.
00:51:42.640 --> 00:51:43.790
If you want a more theoretical
00:51:43.790 --> 00:51:46.070
background, 446 would be more
00:51:46.070 --> 00:51:47.680
specifically for you.
00:51:47.680 --> 00:51:49.590
I think there could be value in taking
00:51:49.590 --> 00:51:50.300
both of them.
00:51:50.300 --> 00:51:52.630
When was I took like a big variety of
00:51:52.630 --> 00:51:55.000
machine learning courses and I never
00:51:55.000 --> 00:51:56.400
really minded if they had overlap
00:51:56.400 --> 00:51:57.770
because I always found it useful to
00:51:57.770 --> 00:51:59.650
revisit a topic and to get different
00:51:59.650 --> 00:52:00.990
perspectives on it.
00:52:00.990 --> 00:52:03.460
But 446 is more theory oriented.
00:52:04.610 --> 00:52:06.770
And if you want to focus on a single
00:52:06.770 --> 00:52:08.460
Application Domain, again there's more
00:52:08.460 --> 00:52:10.680
focused courses for that, like a Vision
00:52:10.680 --> 00:52:13.020
or NLP or special topics course.
00:52:13.680 --> 00:52:15.650
And if you want an easy A, this is also
00:52:15.650 --> 00:52:18.160
probably not the right course.
00:52:18.160 --> 00:52:20.022
It is possible for everybody to get in
00:52:20.022 --> 00:52:21.790
a if you were to.
00:52:22.180 --> 00:52:23.970
To do well on the Assignments and
00:52:23.970 --> 00:52:26.750
exams, but it's definitely not going to
00:52:26.750 --> 00:52:27.210
be easy.
00:52:29.920 --> 00:52:32.090
And as I mentioned earlier, your
00:52:32.090 --> 00:52:33.240
feedback is welcome.
00:52:33.240 --> 00:52:35.715
So I will occasionally solicit feedback
00:52:35.715 --> 00:52:37.590
and if you respond, that would help me.
00:52:38.580 --> 00:52:40.005
You can always talk to me after class
00:52:40.005 --> 00:52:41.860
or send me a message on CampusWire or
00:52:41.860 --> 00:52:43.530
come to my office hours.
00:52:43.630 --> 00:52:47.530
And my philosophy in teaching is to be
00:52:47.530 --> 00:52:48.715
a force multiplier.
00:52:48.715 --> 00:52:51.328
So for every I want every hour of
00:52:51.328 --> 00:52:53.039
effort that you put into the course to
00:52:53.040 --> 00:52:55.150
produce as much learning as possible.
00:52:55.150 --> 00:52:58.090
So it means that you do need to put an
00:52:58.090 --> 00:53:00.419
effort to get value out of the course,
00:53:00.420 --> 00:53:03.040
but I try to select the topics and
00:53:03.040 --> 00:53:03.800
Assignments.
00:53:04.920 --> 00:53:07.420
And organize the materials in a way
00:53:07.420 --> 00:53:09.450
that makes your Learning efficient.
00:53:11.860 --> 00:53:13.710
Alright, so now what do you do next?
00:53:13.710 --> 00:53:16.120
Bookmark the website so that you can go
00:53:16.120 --> 00:53:17.600
back to it easily?
00:53:17.600 --> 00:53:19.160
Sign up for CampusWire?
00:53:19.880 --> 00:53:22.300
Read the syllabus and the schedule and
00:53:22.300 --> 00:53:24.330
I would recommend watching the
00:53:24.330 --> 00:53:25.000
tutorials.
00:53:25.760 --> 00:53:27.650
And then in the next class I'm going to
00:53:27.650 --> 00:53:30.570
talk about K nearest neighbor and kind
00:53:30.570 --> 00:53:32.740
of an overview of classification and
00:53:32.740 --> 00:53:33.710
regression.
00:53:33.710 --> 00:53:35.843
And I'll also introduce homework 1 S
00:53:35.843 --> 00:53:37.400
you can check out homework one if you
00:53:37.400 --> 00:53:40.059
want, or just wait till the next class
00:53:40.060 --> 00:53:41.110
when I introduce it.
00:53:42.990 --> 00:53:45.040
Alright, so that's it for today.
00:53:45.040 --> 00:53:48.420
I'll first before you get up, does
00:53:48.420 --> 00:53:50.000
anybody have any questions that you
00:53:50.000 --> 00:53:51.420
want to ask?
00:53:51.420 --> 00:53:54.460
Wait, don't get up, don't pick up your
00:53:54.460 --> 00:53:55.120
things yet.
00:53:55.120 --> 00:53:55.960
Yes.
00:53:56.240 --> 00:53:56.680
Join the.
00:53:58.240 --> 00:54:00.240
OK.
00:54:00.240 --> 00:54:01.000
Anything else?
00:54:01.780 --> 00:54:03.560
Alright, so just come up to me if you
00:54:03.560 --> 00:54:05.730
have any questions and I'll be here for
00:54:05.730 --> 00:54:06.070
a while.
00:54:07.630 --> 00:54:08.600
See you Thursday.
00:54:10.050 --> 00:54:12.390
A limited submission on the exams.
00:54:12.390 --> 00:54:15.270
What do you have like?
01:14:27.910 --> 01:14:29.830
Testing, testing, testing.
01:14:40.680 --> 01:14:41.020
OK.
01:14:43.690 --> 01:14:44.490
Every time you come.