whisper-finetuning-for-asee / CS_441_2023_Spring_January_31,_2023.vtt
classen3's picture
Imported CS 441 audio/transcripts
a67be9a verified
WEBVTT Kind: captions; Language: en-US
NOTE
Created on 2024-02-07T20:53:13.8059397Z by ClassTranscribe
00:01:57.930 --> 00:01:59.190
It seems like there's like.
00:02:01.950 --> 00:02:02.550
Yes, it's OK.
00:02:03.590 --> 00:02:04.800
Alright, good morning everybody.
00:02:08.160 --> 00:02:10.626
So I thought it I was trying to figure
00:02:10.626 --> 00:02:12.030
out why this seems like there's a lot
00:02:12.030 --> 00:02:13.520
of light on the screen, but I can't
00:02:13.520 --> 00:02:14.192
figure it out.
00:02:14.192 --> 00:02:16.300
I thought it was interesting that this
00:02:16.300 --> 00:02:17.490
for this picture.
00:02:17.580 --> 00:02:18.110
And.
00:02:19.140 --> 00:02:20.760
So I'm generating all of these with
00:02:20.760 --> 00:02:21.510
Dolly.
00:02:21.510 --> 00:02:22.880
This one was a dirt Rd.
00:02:22.880 --> 00:02:24.500
splits around a large gnarly tree
00:02:24.500 --> 00:02:25.640
fractal art.
00:02:25.640 --> 00:02:27.830
But I thought it was really funny how
00:02:27.830 --> 00:02:30.780
it without my bidding it put like some
00:02:30.780 --> 00:02:32.530
kind of superhero or something behind
00:02:32.530 --> 00:02:34.756
the tree there's like some looks like
00:02:34.756 --> 00:02:36.360
there's like some superhero that's like
00:02:36.360 --> 00:02:37.230
flying in and.
00:02:38.130 --> 00:02:39.420
I don't know where that came from.
00:02:41.170 --> 00:02:42.430
Can you guys see the screen OK?
00:02:43.750 --> 00:02:44.830
Seems a little faded.
00:03:05.390 --> 00:03:05.650
But.
00:03:12.610 --> 00:03:13.580
OK I.
00:03:16.440 --> 00:03:18.730
Yeah, I put the lights are on all off.
00:03:21.400 --> 00:03:22.580
But those are still on.
00:03:25.090 --> 00:03:27.020
Alright, let me just take one second.
00:03:48.110 --> 00:03:48.600
All right.
00:03:48.600 --> 00:03:50.860
Anyway, I'll move with it.
00:03:51.500 --> 00:03:53.410
Alright, so.
00:03:53.540 --> 00:03:55.570
And so for some Logistics, I wanted to
00:03:55.570 --> 00:03:57.160
I never got to introduce some of the
00:03:57.160 --> 00:03:58.740
TAS because the couple couldn't be here
00:03:58.740 --> 00:04:00.110
in the first day and I kept forgetting.
00:04:01.130 --> 00:04:03.890
So, Josh, are you here?
00:04:04.950 --> 00:04:08.900
OK, cool if you want to just actually.
00:04:10.130 --> 00:04:11.160
I can give my mic.
00:04:11.160 --> 00:04:12.780
If you want to just introduce yourself
00:04:12.780 --> 00:04:14.540
a little bit, you can say like what
00:04:14.540 --> 00:04:14.940
kind of.
00:04:19.890 --> 00:04:20.450
Yeah.
00:04:20.450 --> 00:04:20.870
Hi, everyone.
00:04:20.870 --> 00:04:21.290
I'm Josh.
00:04:21.290 --> 00:04:23.110
I've been applying machine learning to
00:04:23.110 --> 00:04:25.980
autonomous cars and airplanes.
00:04:27.150 --> 00:04:27.480
Cool.
00:04:27.480 --> 00:04:27.900
Thank you.
00:04:28.830 --> 00:04:31.020
And cassette, cassette.
00:04:37.760 --> 00:04:38.320
Yeah.
00:04:41.120 --> 00:04:46.150
OK hey everyone, I'm a TA for CS441 and
00:04:46.150 --> 00:04:49.230
I have experience with NLP majorly.
00:04:49.230 --> 00:04:49.720
Thank you.
00:04:50.960 --> 00:04:51.190
Great.
00:04:51.190 --> 00:04:52.080
Thank you.
00:04:52.080 --> 00:04:54.230
And I don't think Peter's here, but
00:04:54.230 --> 00:04:55.820
Peter, are you here, OK.
00:04:56.520 --> 00:04:58.240
Usually has a conflict on Tuesday, so
00:04:58.240 --> 00:05:00.780
also we have Pedro is a.
00:05:01.510 --> 00:05:04.170
A pro stock course Assistant.
00:05:04.170 --> 00:05:07.020
So it's not like a regular TA, but
00:05:07.020 --> 00:05:07.650
he's.
00:05:08.730 --> 00:05:10.510
Doing a postdoc with Nancy Amato.
00:05:11.170 --> 00:05:12.880
And here is helping out with the online
00:05:12.880 --> 00:05:15.190
course for a couple semesters.
00:05:15.830 --> 00:05:17.060
And so he's helping out with this
00:05:17.060 --> 00:05:19.140
course and he's.
00:05:20.920 --> 00:05:23.700
One of the things he's doing is holding
00:05:23.700 --> 00:05:25.640
office hours, and so especially if you
00:05:25.640 --> 00:05:27.710
have, if you want help with your
00:05:27.710 --> 00:05:31.940
projects or homeworks, they're like
00:05:31.940 --> 00:05:34.000
higher level advice, then he can be a
00:05:34.000 --> 00:05:35.676
really good resource for that.
00:05:35.676 --> 00:05:37.400
So I know a lot of people want to meet
00:05:37.400 --> 00:05:39.250
with me about their side projects,
00:05:39.250 --> 00:05:40.720
which is also fine, you're welcome to
00:05:40.720 --> 00:05:42.100
do that.
00:05:42.100 --> 00:05:44.090
But he's also a good person for that.
00:05:46.480 --> 00:05:49.550
Alright, so just as a reminder for
00:05:49.550 --> 00:05:51.210
anybody who wasn't here, the first
00:05:51.210 --> 00:05:53.630
lecture, all the notes and everything
00:05:53.630 --> 00:05:55.635
are on this web page.
00:05:55.635 --> 00:05:57.890
So make sure that you go there and sign
00:05:57.890 --> 00:06:00.290
up for CampusWire where announcements
00:06:00.290 --> 00:06:01.600
will be made.
00:06:01.600 --> 00:06:06.120
Also, I sent a survey by e-mail and I
00:06:06.120 --> 00:06:07.760
got a little bit of responses last
00:06:07.760 --> 00:06:08.240
night.
00:06:08.240 --> 00:06:10.390
Do you take some time to respond to it
00:06:10.390 --> 00:06:11.300
please?
00:06:11.300 --> 00:06:12.290
There's two parts.
00:06:12.290 --> 00:06:14.340
One is just asking for feedback about
00:06:14.340 --> 00:06:15.300
like piece of the course.
00:06:15.380 --> 00:06:16.340
And stuff like that.
00:06:16.420 --> 00:06:16.900
And.
00:06:17.720 --> 00:06:20.640
One part is asking about your interests
00:06:20.640 --> 00:06:23.060
for some of the possible.
00:06:24.070 --> 00:06:26.332
Challenges that I'll pick for final
00:06:26.332 --> 00:06:29.810
project and so basically for the final
00:06:29.810 --> 00:06:32.149
project there will be 3 challenges that
00:06:32.150 --> 00:06:33.600
are like pre selected.
00:06:34.230 --> 00:06:35.720
But if you don't want to do those, you
00:06:35.720 --> 00:06:38.370
can also just do some benchmark that's
00:06:38.370 --> 00:06:40.070
online or you can even do a custom
00:06:40.070 --> 00:06:40.860
task.
00:06:40.860 --> 00:06:43.960
And I'll post the specifications for
00:06:43.960 --> 00:06:46.640
final project soon as homework 2.
00:06:47.960 --> 00:06:50.140
Also, just based on the feedback I've
00:06:50.140 --> 00:06:52.810
seen so far, I think nobody thinks it's
00:06:52.810 --> 00:06:54.570
way too easy or too slow.
00:06:54.570 --> 00:06:57.150
Some people think it's much too fast
00:06:57.150 --> 00:06:57.930
and too hard.
00:06:57.930 --> 00:06:59.710
So I'm going to take some time on
00:06:59.710 --> 00:07:03.450
Thursday to Reconsolidate and present.
00:07:03.450 --> 00:07:07.280
Kind of go over what we've done so far,
00:07:07.280 --> 00:07:09.750
talk in more depth or maybe not more
00:07:09.750 --> 00:07:11.439
depth, but at least go over the
00:07:11.440 --> 00:07:12.150
concepts.
00:07:13.270 --> 00:07:16.080
And the algorithms and a little bit of
00:07:16.080 --> 00:07:18.380
code now that you've had a first pass
00:07:18.380 --> 00:07:18.640
edit.
00:07:20.460 --> 00:07:22.830
So I'll tap the brakes a little bit to
00:07:22.830 --> 00:07:24.500
do that because I think it's really
00:07:24.500 --> 00:07:27.215
important that these that everyone is
00:07:27.215 --> 00:07:28.790
really solid on these fundamentals.
00:07:28.790 --> 00:07:31.260
And I know that there's a pretty big
00:07:31.260 --> 00:07:33.090
range of backgrounds of people taking
00:07:33.090 --> 00:07:35.060
the course, many people from other
00:07:35.060 --> 00:07:35.710
departments.
00:07:37.290 --> 00:07:39.900
As well as other different kinds of.
00:07:41.230 --> 00:07:43.280
Of like academic foundations.
00:07:44.270 --> 00:07:44.610
Alright.
00:07:45.910 --> 00:07:47.890
So just to recap what we talked about
00:07:47.890 --> 00:07:49.640
in the last few lectures, very briefly,
00:07:49.640 --> 00:07:51.040
we talked about Nearest neighbor.
00:07:51.780 --> 00:07:53.210
And the superpower is the nearest
00:07:53.210 --> 00:07:55.170
neighbor are that it can instantly
00:07:55.170 --> 00:07:56.230
learn new classes.
00:07:56.230 --> 00:07:58.020
You can just add a new example to your
00:07:58.020 --> 00:07:58.790
training set.
00:07:58.790 --> 00:08:00.780
And since there's no model that has to
00:08:00.780 --> 00:08:04.110
be like tuned, you can just learn super
00:08:04.110 --> 00:08:04.720
quickly.
00:08:04.720 --> 00:08:07.450
And it's also a pretty good predictor
00:08:07.450 --> 00:08:08.980
from either one or many examples.
00:08:08.980 --> 00:08:10.430
So it's a really good.
00:08:10.530 --> 00:08:13.690
It's a really good algorithm to have in
00:08:13.690 --> 00:08:15.330
your tool belt as a baseline and
00:08:15.330 --> 00:08:16.760
sometimes as a best performer.
00:08:18.500 --> 00:08:20.160
We also talked about Naive bees.
00:08:21.050 --> 00:08:24.140
Night Bayes is not a great performer as
00:08:24.140 --> 00:08:26.984
like a full algorithm, but it's often
00:08:26.984 --> 00:08:27.426
a.
00:08:27.426 --> 00:08:30.075
It's an important concept because it's
00:08:30.075 --> 00:08:31.760
often part of an assumption that you
00:08:31.760 --> 00:08:32.920
make when you're trying to model
00:08:32.920 --> 00:08:35.560
probabilities that you'll assume that
00:08:35.560 --> 00:08:37.630
the different features are independent
00:08:37.630 --> 00:08:39.010
given the thing that you're trying to
00:08:39.010 --> 00:08:39.330
predict.
00:08:41.780 --> 00:08:44.290
It does have its pros, so the pros are
00:08:44.290 --> 00:08:46.560
that it's really fast to estimate even
00:08:46.560 --> 00:08:48.113
if you've got a lot of data.
00:08:48.113 --> 00:08:49.909
And if you don't have a lot of data and
00:08:49.910 --> 00:08:51.130
you're trying to get a probabilistic
00:08:51.130 --> 00:08:53.300
classifier, then it might be your best
00:08:53.300 --> 00:08:53.750
choice.
00:08:53.750 --> 00:08:56.700
Because of its strong assumptions, you
00:08:56.700 --> 00:08:59.880
can get decent estimates on those
00:08:59.880 --> 00:09:02.160
single variable functions from even
00:09:02.160 --> 00:09:02.800
limited data.
00:09:05.460 --> 00:09:07.964
We talked about logistic regression.
00:09:07.964 --> 00:09:10.830
Logistic regression is another super
00:09:10.830 --> 00:09:12.230
widely used classifier.
00:09:13.580 --> 00:09:16.400
I think the AML book says that SVM is
00:09:16.400 --> 00:09:18.776
should be or like go to as a first as
00:09:18.776 --> 00:09:20.660
like a first thing you try, but in my
00:09:20.660 --> 00:09:22.130
opinion Logistic Regression is.
00:09:23.810 --> 00:09:25.085
It's very effective.
00:09:25.085 --> 00:09:26.810
It's a very effective predictor if you
00:09:26.810 --> 00:09:28.720
have high dimensional features and it
00:09:28.720 --> 00:09:30.250
also provides good confidence
00:09:30.250 --> 00:09:31.460
estimates, meaning that.
00:09:32.150 --> 00:09:35.320
You get not only most likely class, but
00:09:35.320 --> 00:09:37.470
the probability that prediction is
00:09:37.470 --> 00:09:40.800
correct and those probabilities fairly
00:09:40.800 --> 00:09:41.400
trustworthy.
00:09:43.320 --> 00:09:44.970
We also talked about Linear Regression,
00:09:44.970 --> 00:09:46.560
where you're fitting a line to a set of
00:09:46.560 --> 00:09:47.050
points.
00:09:47.670 --> 00:09:50.610
And you can extrapolate to predict like
00:09:50.610 --> 00:09:52.620
new values that are outside of your
00:09:52.620 --> 00:09:53.790
Training range.
00:09:54.530 --> 00:09:55.840
And so.
00:09:56.730 --> 00:09:58.450
Linear regression is also useful for
00:09:58.450 --> 00:10:00.270
explaining relationships you're very
00:10:00.270 --> 00:10:02.100
commonly see, like trend lines.
00:10:02.100 --> 00:10:03.390
That's just Linear Regression.
00:10:04.130 --> 00:10:05.850
And you can predict continuous values
00:10:05.850 --> 00:10:07.600
from many variables in linear
00:10:07.600 --> 00:10:10.130
regression is also like probably the
00:10:10.130 --> 00:10:12.760
most common tool for.
00:10:12.830 --> 00:10:15.590
For things like, I don't know, like
00:10:15.590 --> 00:10:17.790
economics or analyzing.
00:10:18.770 --> 00:10:23.100
Yeah, time series analyzing like fMRI
00:10:23.100 --> 00:10:25.930
data or all kinds of scientific and
00:10:25.930 --> 00:10:27.180
economic analysis.
00:10:30.420 --> 00:10:33.810
So almost all algorithms involve these
00:10:33.810 --> 00:10:35.760
Nearest neighbor, logistic regression
00:10:35.760 --> 00:10:36.850
or linear regression.
00:10:37.540 --> 00:10:41.040
And the reason that there's thousand
00:10:41.040 --> 00:10:43.330
papers published in the last 10 years
00:10:43.330 --> 00:10:45.060
or so, probably a lot more than that
00:10:45.060 --> 00:10:47.030
actually, is that.
00:10:47.850 --> 00:10:50.120
Is really the feature learning, so it's
00:10:50.120 --> 00:10:52.090
getting the right representation so
00:10:52.090 --> 00:10:54.490
that when you feed that representation
00:10:54.490 --> 00:10:56.610
into these like Linear models or
00:10:56.610 --> 00:10:59.080
Nearest neighbor, you get good results.
00:11:00.080 --> 00:11:00.660
And so.
00:11:01.510 --> 00:11:03.020
Pretty much the rest of what we're
00:11:03.020 --> 00:11:05.160
going to learn in the supervised
00:11:05.160 --> 00:11:07.520
learning section of the course is how
00:11:07.520 --> 00:11:08.460
to learn features.
00:11:11.930 --> 00:11:14.150
So I did want to just briefly go over
00:11:14.150 --> 00:11:15.640
the homework and remind you that it's
00:11:15.640 --> 00:11:18.180
due on February 6th on Monday.
00:11:19.060 --> 00:11:21.400
And I'll be going over some related
00:11:21.400 --> 00:11:22.830
things again in more detail on
00:11:22.830 --> 00:11:23.350
Thursday.
00:11:24.300 --> 00:11:27.200
But there's two parts to the main
00:11:27.200 --> 00:11:27.590
homework.
00:11:27.590 --> 00:11:29.770
There's Digit Classification where
00:11:29.770 --> 00:11:31.186
you're trying to predict a label zero
00:11:31.186 --> 00:11:33.530
to 9 based on a 28 by 28 image.
00:11:34.410 --> 00:11:36.409
These images get reshaped into like a
00:11:36.410 --> 00:11:38.840
single vector, so you have a feature
00:11:38.840 --> 00:11:41.020
vector that corresponds to the pixel
00:11:41.020 --> 00:11:42.280
intensities of the image.
00:11:44.350 --> 00:11:46.510
And then you have to do K and Naive
00:11:46.510 --> 00:11:49.200
Bayes, linear logistic regression.
00:11:50.060 --> 00:11:52.510
And plot the Error versus.
00:11:52.670 --> 00:11:55.420
A plot Error versus Training size to
00:11:55.420 --> 00:11:57.310
get a sense for like how performance
00:11:57.310 --> 00:11:58.745
changes as you vary the number of
00:11:58.745 --> 00:11:59.530
training examples.
00:12:00.380 --> 00:12:02.820
And then to select the best parameter
00:12:02.820 --> 00:12:06.490
using validation set which is a really
00:12:06.490 --> 00:12:07.240
hyper parameter.
00:12:07.240 --> 00:12:09.020
Tuning is like something that you do
00:12:09.020 --> 00:12:10.100
all the time in machine learning.
00:12:13.150 --> 00:12:14.839
The second problem is Temperature
00:12:14.840 --> 00:12:15.700
Regression.
00:12:15.700 --> 00:12:18.182
So I got this Temperature.
00:12:18.182 --> 00:12:20.178
This data set of like the temperature
00:12:20.178 --> 00:12:22.890
is a big cities in the US and then
00:12:22.890 --> 00:12:24.111
made-up a problem from it.
00:12:24.111 --> 00:12:26.550
So the problem is to try to predict the
00:12:26.550 --> 00:12:28.220
next day's temperature in Cleveland
00:12:28.220 --> 00:12:30.000
which stays zero given all the previous
00:12:30.000 --> 00:12:30.550
temperatures.
00:12:31.490 --> 00:12:34.350
And these features have meanings.
00:12:34.350 --> 00:12:37.417
Every feature is some previous is, like
00:12:37.417 --> 00:12:39.530
the temperature of 1 in the big cities
00:12:39.530 --> 00:12:40.990
from one in the past five days.
00:12:42.570 --> 00:12:44.753
But you can kind of.
00:12:44.753 --> 00:12:46.110
You don't really need to know those
00:12:46.110 --> 00:12:48.110
meanings in order to solve the problem
00:12:48.110 --> 00:12:48.480
again.
00:12:48.480 --> 00:12:50.780
You essentially just have a feature
00:12:50.780 --> 00:12:53.730
vector of a bunch of continuous values
00:12:53.730 --> 00:12:55.570
in this case, and you're trying to
00:12:55.570 --> 00:12:57.290
predict a new continuous value, which
00:12:57.290 --> 00:13:00.460
is Cleveland Cleveland's temperature in
00:13:00.460 --> 00:13:01.240
the next day.
00:13:02.010 --> 00:13:04.545
And again you can use KNN and a Bayes
00:13:04.545 --> 00:13:06.000
and now Linear Regression.
00:13:07.020 --> 00:13:08.935
KNN implementation will be essentially
00:13:08.935 --> 00:13:11.440
the same for these A2 line change of
00:13:11.440 --> 00:13:13.820
code because now instead of predicting
00:13:13.820 --> 00:13:16.230
a categorical variable, you're
00:13:16.230 --> 00:13:18.440
predicting a continuous variable.
00:13:18.440 --> 00:13:20.580
So if K is greater than one, you
00:13:20.580 --> 00:13:23.770
average the predictions for Regression
00:13:23.770 --> 00:13:26.280
where for the Classification you choose
00:13:26.280 --> 00:13:27.430
the most common prediction.
00:13:28.580 --> 00:13:29.840
That's the only change.
00:13:29.840 --> 00:13:31.590
Naive Bayes does change quite a bit
00:13:31.590 --> 00:13:32.710
because you're using a different
00:13:32.710 --> 00:13:33.590
probabilistic model.
00:13:34.360 --> 00:13:36.710
And remember that there's one lecture
00:13:36.710 --> 00:13:38.670
slide that has the derivation for how
00:13:38.670 --> 00:13:40.545
you do the inference for nibbies under
00:13:40.545 --> 00:13:41.050
the setting.
00:13:42.330 --> 00:13:44.760
And then for linear and logistic
00:13:44.760 --> 00:13:46.820
regression you're able to use the
00:13:46.820 --> 00:13:48.350
modules in sklearn.
00:13:49.890 --> 00:13:51.680
And then the final part is to identify
00:13:51.680 --> 00:13:53.550
the most important features using L1
00:13:53.550 --> 00:13:54.320
Linear Regression.
00:13:55.030 --> 00:13:57.160
So the reason that we use.
00:13:58.020 --> 00:13:59.810
And when we do like.
00:14:01.000 --> 00:14:03.170
Linear and logistic regression, we're
00:14:03.170 --> 00:14:03.580
trying.
00:14:03.580 --> 00:14:05.228
We're mainly trying to fit the data.
00:14:05.228 --> 00:14:06.600
We're trying to come up with a model
00:14:06.600 --> 00:14:08.340
that fits the data or fits our
00:14:08.340 --> 00:14:09.920
predictions given the features.
00:14:10.630 --> 00:14:13.720
But also we often express some
00:14:13.720 --> 00:14:14.490
preference.
00:14:15.190 --> 00:14:19.892
Over the model, in particular that the
00:14:19.892 --> 00:14:21.669
weights don't get too large, and the
00:14:21.670 --> 00:14:25.170
reason for that is to avoid like
00:14:25.170 --> 00:14:27.070
overfitting or over relying on
00:14:27.070 --> 00:14:30.410
particular features, as well as to
00:14:30.410 --> 00:14:34.795
improve the generalization to new data
00:14:34.795 --> 00:14:36.209
and generalization.
00:14:36.210 --> 00:14:37.810
Research shows that if you can fit
00:14:37.810 --> 00:14:39.220
something with smaller weights, then
00:14:39.220 --> 00:14:42.013
you're more likely to generalize to new
00:14:42.013 --> 00:14:42.209
data.
00:14:44.640 --> 00:14:46.550
And here we're going to use it for
00:14:46.550 --> 00:14:47.520
feature selection, yeah?
00:14:58.910 --> 00:15:02.370
The so the parameters are.
00:15:02.370 --> 00:15:04.510
You're talking about 1/3.
00:15:04.510 --> 00:15:07.825
OK, so for Naive Bayes the parameter is
00:15:07.825 --> 00:15:10.430
the prior, so that's like the alpha of
00:15:10.430 --> 00:15:11.010
like your.
00:15:11.670 --> 00:15:14.903
In the, it's the initial count, so you
00:15:14.903 --> 00:15:15.882
have a Naive Bayes.
00:15:15.882 --> 00:15:17.360
You have a prior that's essentially
00:15:17.360 --> 00:15:19.200
that you pretend like you've seen all
00:15:19.200 --> 00:15:20.230
combinations of.
00:15:20.950 --> 00:15:23.930
Of things that you're counting, you
00:15:23.930 --> 00:15:26.210
pretend that you see alpha times, and
00:15:26.210 --> 00:15:28.510
so that kind of gives you a bias
00:15:28.510 --> 00:15:30.200
towards estimating that everything's
00:15:30.200 --> 00:15:33.170
equally likely, and that alpha is a
00:15:33.170 --> 00:15:34.190
parameter that you can use.
00:15:34.810 --> 00:15:36.270
You can learn using validation.
00:15:37.010 --> 00:15:39.920
For Logistic Regression, it's the
00:15:39.920 --> 00:15:42.650
Lambda which is your weight on the
00:15:42.650 --> 00:15:43.760
regularization term.
00:15:45.650 --> 00:15:48.180
And for K&N, it's your K, which is the
00:15:48.180 --> 00:15:49.320
number of nearest neighbors you
00:15:49.320 --> 00:15:49.710
consider.
00:15:57.960 --> 00:15:58.220
Yeah.
00:16:00.180 --> 00:16:03.284
So the K&N is.
00:16:03.284 --> 00:16:05.686
It's almost the same whether you're
00:16:05.686 --> 00:16:08.260
doing Regression or Classification.
00:16:08.260 --> 00:16:09.980
When you find the K nearest neighbors,
00:16:09.980 --> 00:16:11.790
it's the exact same code.
00:16:11.790 --> 00:16:14.016
The difference is that if you're doing
00:16:14.016 --> 00:16:15.270
Regression, you're trying to predict
00:16:15.270 --> 00:16:16.200
continuous values.
00:16:16.200 --> 00:16:19.340
So if K is greater than one, then you
00:16:19.340 --> 00:16:21.532
want to average those continuous values
00:16:21.532 --> 00:16:23.150
to get your final prediction.
00:16:23.850 --> 00:16:26.060
And if you're doing Classification, you
00:16:26.060 --> 00:16:28.490
find the most common label instead of
00:16:28.490 --> 00:16:29.803
averaging because you don't want to
00:16:29.803 --> 00:16:31.470
say, well it could be a four, it could
00:16:31.470 --> 00:16:31.980
be a 9.
00:16:31.980 --> 00:16:33.110
So I'm going to like split the
00:16:33.110 --> 00:16:34.270
difference and say it's a 6.
00:16:42.030 --> 00:16:45.420
That are averaging just that you so
00:16:45.420 --> 00:16:47.600
like if K&N returns like the
00:16:47.600 --> 00:16:54.870
temperatures of 1012 and 13 then you
00:16:54.870 --> 00:16:57.190
would say that the average temperature
00:16:57.190 --> 00:16:59.530
is like 11.3 or whatever that works out
00:16:59.530 --> 00:16:59.730
to.
00:17:04.600 --> 00:17:06.440
Yeah, at the end, if K is greater than
00:17:06.440 --> 00:17:09.333
one, then you take the arithmetic mean
00:17:09.333 --> 00:17:11.210
of the average of the.
00:17:11.940 --> 00:17:14.560
Predictions of your K nearest
00:17:14.560 --> 00:17:14.970
neighbors.
00:17:16.590 --> 00:17:16.790
Yeah.
00:17:18.610 --> 00:17:20.560
And so you could also get a variance
00:17:20.560 --> 00:17:22.370
from that, which you don't need to do
00:17:22.370 --> 00:17:24.500
for the homework, but so as a result
00:17:24.500 --> 00:17:26.550
you can have some like confidence bound
00:17:26.550 --> 00:17:27.840
on your estimate as well.
00:17:30.050 --> 00:17:31.780
Alright then you have stretch goals,
00:17:31.780 --> 00:17:32.170
so.
00:17:32.850 --> 00:17:34.100
Stretch goals are.
00:17:35.130 --> 00:17:37.000
Mainly intended for people taking the
00:17:37.000 --> 00:17:39.020
four credit version, but you can anyone
00:17:39.020 --> 00:17:39.510
can try them.
00:17:40.240 --> 00:17:42.570
So there's just improving the MNIST
00:17:42.570 --> 00:17:44.370
classification, like some ideas.
00:17:44.370 --> 00:17:47.360
Or you could try to crop around the
00:17:47.360 --> 00:17:49.000
Digit, or you could make sure that
00:17:49.000 --> 00:17:51.840
they're all centered, or do some
00:17:51.840 --> 00:17:53.410
whitening or other kinds of feature
00:17:53.410 --> 00:17:54.340
transformations.
00:17:55.430 --> 00:17:56.770
Improving Temperature Regression.
00:17:56.770 --> 00:18:00.070
To be honest, I'm not sure exactly how
00:18:00.070 --> 00:18:01.829
much this can be improved or how to
00:18:01.830 --> 00:18:02.280
improve it.
00:18:03.030 --> 00:18:04.720
Again, there's.
00:18:04.720 --> 00:18:07.370
What I would do is try like subtracting
00:18:07.370 --> 00:18:08.110
off the mean.
00:18:08.110 --> 00:18:09.220
For example, you can.
00:18:10.380 --> 00:18:12.370
You can normalize your features before
00:18:12.370 --> 00:18:15.540
you do the fitting by subtracting off
00:18:15.540 --> 00:18:16.750
means and dividing by steering
00:18:16.750 --> 00:18:17.410
deviations.
00:18:17.410 --> 00:18:18.140
That's one idea.
00:18:19.060 --> 00:18:22.320
But we'll look at it after submissions
00:18:22.320 --> 00:18:24.095
if it turns out that.
00:18:24.095 --> 00:18:27.020
So the targets I Choose are because I
00:18:27.020 --> 00:18:29.273
was able to do like some simple things
00:18:29.273 --> 00:18:32.383
to bring down the Error by a few tenths
00:18:32.383 --> 00:18:33.510
of a percent.
00:18:33.510 --> 00:18:35.000
So I kind of figured that if you do
00:18:35.000 --> 00:18:36.346
more things, you'll be able to bring it
00:18:36.346 --> 00:18:38.420
down further, but it's hard to tell so.
00:18:39.240 --> 00:18:40.960
If you do this and you put a lot of
00:18:40.960 --> 00:18:42.600
effort into it, describe your effort
00:18:42.600 --> 00:18:45.594
and we'll assign points even if you
00:18:45.594 --> 00:18:47.680
even if it turns out that there's not
00:18:47.680 --> 00:18:48.640
like a big improvement.
00:18:48.640 --> 00:18:50.676
So don't stress out if you can't get
00:18:50.676 --> 00:18:51.609
like 119.
00:18:52.450 --> 00:18:54.200
RMS a year or something like that.
00:18:55.130 --> 00:18:55.335
Right.
00:18:55.335 --> 00:18:57.306
The last one is to generate a train
00:18:57.306 --> 00:18:58.806
set, train Test, Classification set.
00:18:58.806 --> 00:19:00.380
So this actually means don't like
00:19:00.380 --> 00:19:02.020
generate it out of MNIST to create
00:19:02.020 --> 00:19:02.804
synthetic data.
00:19:02.804 --> 00:19:05.020
So you can Naive Bayes make certain
00:19:05.020 --> 00:19:05.405
assumptions.
00:19:05.405 --> 00:19:07.180
So if you generate your data according
00:19:07.180 --> 00:19:09.390
to those Assumptions, you should be
00:19:09.390 --> 00:19:11.900
able to create a problem that we're
00:19:11.900 --> 00:19:13.520
Naive bees can outperform the other
00:19:13.520 --> 00:19:13.980
methods.
00:19:18.970 --> 00:19:22.130
So for these homeworks, make sure that
00:19:22.130 --> 00:19:24.020
you of course read the assignment.
00:19:24.020 --> 00:19:25.040
Read the tips.
00:19:25.530 --> 00:19:26.210
00:19:27.060 --> 00:19:29.190
And then you should be adding code to
00:19:29.190 --> 00:19:30.045
the starter code.
00:19:30.045 --> 00:19:31.610
The starter code doesn't really solve
00:19:31.610 --> 00:19:33.030
the problems for you, but it loads the
00:19:33.030 --> 00:19:34.570
data and gives you some examples.
00:19:34.570 --> 00:19:38.160
So for example, for example, there's a.
00:19:38.810 --> 00:19:41.340
In the Regression, I think it includes
00:19:41.340 --> 00:19:43.710
like a baseline where it computes RMSE
00:19:43.710 --> 00:19:46.450
and median absolute error, so that
00:19:46.450 --> 00:19:48.760
function can essentially be reused
00:19:48.760 --> 00:19:50.060
later to compute the errors.
00:19:51.120 --> 00:19:53.073
And that baseline gives you some idea
00:19:53.073 --> 00:19:55.390
of like what kind of performance you
00:19:55.390 --> 00:19:55.870
might get.
00:19:55.870 --> 00:19:57.320
Like you should beat that baseline
00:19:57.320 --> 00:19:58.620
because that's just based on a single
00:19:58.620 --> 00:19:58.940
feature.
00:20:00.300 --> 00:20:02.980
And then you complete the report and
00:20:02.980 --> 00:20:04.940
make sure to include expected points.
00:20:04.940 --> 00:20:07.040
So when the grader is graded they will
00:20:07.040 --> 00:20:09.140
essentially just say if they disagree
00:20:09.140 --> 00:20:09.846
with you.
00:20:09.846 --> 00:20:12.470
So you if you claim like 10 points but
00:20:12.470 --> 00:20:14.345
something was wrong then they might say
00:20:14.345 --> 00:20:16.310
you lose like 3 points for this reason
00:20:16.310 --> 00:20:20.340
and so that streamlines their grading.
00:20:21.930 --> 00:20:23.580
The assignment, the report Submit your
00:20:23.580 --> 00:20:26.070
notebook and either if you just have
00:20:26.070 --> 00:20:28.800
one file, submitting the IPYNB is fine
00:20:28.800 --> 00:20:29.890
or otherwise you can zip it.
00:20:30.860 --> 00:20:32.220
And that's it.
00:20:33.960 --> 00:20:34.620
Yeah, question.
00:20:41.730 --> 00:20:47.160
So you need in three Credit was at 450,
00:20:47.160 --> 00:20:47.640
is that right?
00:20:48.650 --> 00:20:50.810
So think I think in the three credit
00:20:50.810 --> 00:20:52.300
you need 450 points.
00:20:53.660 --> 00:20:55.430
Each assignment without doing any
00:20:55.430 --> 00:20:56.230
stretch goals.
00:20:56.230 --> 00:20:58.620
Each assignment is worth 100 points and
00:20:58.620 --> 00:21:01.240
the final project is worth 50 points.
00:21:01.240 --> 00:21:02.673
I mean sorry, the final projects worth
00:21:02.673 --> 00:21:03.460
100 points also.
00:21:04.150 --> 00:21:06.310
So if you're in the three Credit
00:21:06.310 --> 00:21:08.210
version and you don't do any stretch
00:21:08.210 --> 00:21:10.960
goals, and you do all the assignments
00:21:10.960 --> 00:21:12.500
and you do the final project, you will
00:21:12.500 --> 00:21:13.570
have more points than you need.
00:21:14.190 --> 00:21:17.740
So the so you can kind of pick
00:21:17.740 --> 00:21:19.270
something that you don't want to do and
00:21:19.270 --> 00:21:20.910
skip it if you're in the three credit
00:21:20.910 --> 00:21:24.100
course and or like if you just are
00:21:24.100 --> 00:21:26.330
already a machine learning guru, you
00:21:26.330 --> 00:21:29.290
can do like 3 assignments with all the
00:21:29.290 --> 00:21:31.630
extra parts and then take a vacation.
00:21:32.920 --> 00:21:34.720
If you're in the four credit version,
00:21:34.720 --> 00:21:37.490
then you will have to do some of the.
00:21:37.670 --> 00:21:39.520
Some of the stretch goals in order to
00:21:39.520 --> 00:21:41.470
get your full points, which are 550.
00:21:49.580 --> 00:21:52.715
Alright, so now I'm going to move on to
00:21:52.715 --> 00:21:54.180
the main topic.
00:21:54.180 --> 00:21:57.340
So we've seen so far, we've seen 2 main
00:21:57.340 --> 00:21:59.116
choices for how to use the features.
00:21:59.116 --> 00:22:01.025
We could do Nearest neighbor when we
00:22:01.025 --> 00:22:03.200
use all the features jointly in order
00:22:03.200 --> 00:22:05.280
to find similar examples, and then we
00:22:05.280 --> 00:22:06.970
predict the most similar label.
00:22:07.910 --> 00:22:10.160
Or we can use a linear model where
00:22:10.160 --> 00:22:11.980
essentially you're making a prediction
00:22:11.980 --> 00:22:14.530
out of a of all the feature values.
00:22:16.070 --> 00:22:18.490
But there's some other things that are
00:22:18.490 --> 00:22:20.270
kind of intuitive, so.
00:22:21.220 --> 00:22:24.010
For example, if you consider this where
00:22:24.010 --> 00:22:26.260
you're trying to split the red X's from
00:22:26.260 --> 00:22:27.710
the Green O's.
00:22:28.370 --> 00:22:30.820
What's like another way that you might
00:22:30.820 --> 00:22:33.180
try to define what that Decision
00:22:33.180 --> 00:22:35.130
boundary is if you wanted to, say, tell
00:22:35.130 --> 00:22:35.730
somebody else?
00:22:35.730 --> 00:22:37.110
Like how do you identify whether
00:22:37.110 --> 00:22:38.770
something is a no?
00:22:52.240 --> 00:22:55.600
Yeah, I mean you so your jaw some kind
00:22:55.600 --> 00:22:56.200
of boundary.
00:22:57.150 --> 00:22:57.690
And.
00:22:58.620 --> 00:23:00.315
And one way that you might think about
00:23:00.315 --> 00:23:03.440
that is creating a kind of like simple
00:23:03.440 --> 00:23:04.220
rule like this.
00:23:04.220 --> 00:23:05.890
Like you might say that if.
00:23:06.600 --> 00:23:09.040
You basically draw a boundary, but if
00:23:09.040 --> 00:23:11.252
you want to specify you might say if X2
00:23:11.252 --> 00:23:15.820
is less than .6 and X2 is greater than
00:23:15.820 --> 00:23:16.500
two.
00:23:17.460 --> 00:23:21.480
And tX2, oops, that's just say X1 and
00:23:21.480 --> 00:23:22.082
the last one.
00:23:22.082 --> 00:23:24.630
And if X1 is less than seven then it's
00:23:24.630 --> 00:23:26.672
an O and otherwise it's an X.
00:23:26.672 --> 00:23:28.110
So basically you could create like a
00:23:28.110 --> 00:23:29.502
set of rules like that, right?
00:23:29.502 --> 00:23:32.161
So say if it meets these criteria then
00:23:32.161 --> 00:23:34.819
it's one class and if it meets these
00:23:34.820 --> 00:23:37.070
other criteria it's another class.
00:23:40.160 --> 00:23:42.930
And So what we're going to learn today
00:23:42.930 --> 00:23:45.280
is how we can try to learn these rules
00:23:45.280 --> 00:23:48.220
automatically, even if we have a lot of
00:23:48.220 --> 00:23:50.520
features in more complicated kinds of
00:23:50.520 --> 00:23:51.250
predictions.
00:23:52.920 --> 00:23:55.108
So this is basically the idea of
00:23:55.108 --> 00:23:55.744
Decision trees.
00:23:55.744 --> 00:23:58.490
So we all use Decision trees in our own
00:23:58.490 --> 00:24:00.264
life, even if we don't think about it
00:24:00.264 --> 00:24:00.812
that way.
00:24:00.812 --> 00:24:02.710
Like you often say, if this happens,
00:24:02.710 --> 00:24:04.121
I'll do that, and if it doesn't, then
00:24:04.121 --> 00:24:05.029
I'll do this other thing.
00:24:05.030 --> 00:24:06.685
That's like a Decision tree, right?
00:24:06.685 --> 00:24:10.400
You had some kind of criteria, and
00:24:10.400 --> 00:24:12.306
depending on the outcome of that
00:24:12.306 --> 00:24:13.886
criteria, you do one thing.
00:24:13.886 --> 00:24:16.680
And if it's the other way, if you get
00:24:16.680 --> 00:24:17.900
the other outcome, then you would be
00:24:17.900 --> 00:24:18.920
doing the other thing.
00:24:18.920 --> 00:24:20.310
And maybe you have a whole chain of
00:24:20.310 --> 00:24:22.090
them if I.
00:24:22.250 --> 00:24:23.700
If I have time today, I'm going to go
00:24:23.700 --> 00:24:25.990
to the grocery store, but if the car is
00:24:25.990 --> 00:24:27.330
not there then I'm going to do this
00:24:27.330 --> 00:24:28.480
instead and so on.
00:24:29.850 --> 00:24:32.370
All right, so in Decision trees, the
00:24:32.370 --> 00:24:34.500
Training is essentially to iteratively
00:24:34.500 --> 00:24:37.340
Choose the attribute and split in a
00:24:37.340 --> 00:24:40.080
split value that will best separate
00:24:40.080 --> 00:24:41.530
your classes from each other.
00:24:42.920 --> 00:24:44.610
Or if you're doing continuous values
00:24:44.610 --> 00:24:47.010
that kind of group things into similar
00:24:47.010 --> 00:24:48.240
prediction values.
00:24:49.480 --> 00:24:52.440
So for example you might say if these
00:24:52.440 --> 00:24:56.600
red circles are oranges and these
00:24:56.600 --> 00:24:59.264
triangles are lemons, where there
00:24:59.264 --> 00:25:01.090
oranges and lemons are plotted
00:25:01.090 --> 00:25:02.250
according to their width and their
00:25:02.250 --> 00:25:02.750
height.
00:25:02.750 --> 00:25:07.726
You might decide well if it's less than
00:25:07.726 --> 00:25:10.170
6.5 centimeters then.
00:25:10.170 --> 00:25:12.690
Or I'll use greater since it's there if
00:25:12.690 --> 00:25:14.190
it's greater than 6.5 centimeters.
00:25:15.450 --> 00:25:17.267
Then I'm going to split it into this
00:25:17.267 --> 00:25:19.410
section where it's like mostly oranges
00:25:19.410 --> 00:25:22.110
and if it's less than 6.5 centimeters
00:25:22.110 --> 00:25:24.395
width, then I'll split it into this
00:25:24.395 --> 00:25:26.220
section where it's mostly lemons.
00:25:27.250 --> 00:25:30.560
Neither of these a perfect split still.
00:25:30.560 --> 00:25:32.910
So then I go further and say if it was
00:25:32.910 --> 00:25:35.309
on this side of the split, if it's
00:25:35.310 --> 00:25:37.915
greater than 95 centimeter height then
00:25:37.915 --> 00:25:40.350
it's a lemon, and if it's less than
00:25:40.350 --> 00:25:42.130
that then it's a.
00:25:42.820 --> 00:25:43.760
Then it's an orange.
00:25:44.900 --> 00:25:46.660
And now that's like a pretty confident
00:25:46.660 --> 00:25:47.170
prediction.
00:25:47.930 --> 00:25:49.610
And then if I'm on this side then I can
00:25:49.610 --> 00:25:51.560
split it by height and say if it's less
00:25:51.560 --> 00:25:51.990
than.
00:25:53.690 --> 00:25:55.530
If it's greater than 6 centimeters then
00:25:55.530 --> 00:25:57.714
it's a lemon, and if it's less than 6
00:25:57.714 --> 00:25:59.450
centimeters then it's an orange.
00:25:59.450 --> 00:26:01.130
So you can like iteratively Choose a
00:26:01.130 --> 00:26:03.180
test and then keep splitting the data.
00:26:03.780 --> 00:26:06.510
And every time you choose a test, test
00:26:06.510 --> 00:26:09.510
another test that splits the data
00:26:09.510 --> 00:26:10.910
further according to what you're trying
00:26:10.910 --> 00:26:11.320
to predict.
00:26:12.270 --> 00:26:14.890
Essentially, this method Combines a
00:26:14.890 --> 00:26:16.760
feature selection and modeling with
00:26:16.760 --> 00:26:17.410
prediction.
00:26:18.670 --> 00:26:20.420
So at the end of this, you transform
00:26:20.420 --> 00:26:22.940
what we're two continuous values into
00:26:22.940 --> 00:26:24.770
these four discrete values.
00:26:25.450 --> 00:26:27.360
Of different chunks, different
00:26:27.360 --> 00:26:30.130
partitions of the feature space and for
00:26:30.130 --> 00:26:31.350
each of those.
00:26:32.420 --> 00:26:34.850
Each of those parts of the partition.
00:26:35.810 --> 00:26:38.360
You make a prediction.
00:26:39.240 --> 00:26:41.620
A partitioning is just when you take a
00:26:41.620 --> 00:26:44.390
continuous space and divide it up into
00:26:44.390 --> 00:26:46.850
different cells that cover the entire
00:26:46.850 --> 00:26:47.400
space.
00:26:47.400 --> 00:26:49.859
That's a partition where the cells
00:26:49.860 --> 00:26:51.040
don't overlap with each other.
00:26:54.340 --> 00:26:56.460
And then if you want to classify, once
00:26:56.460 --> 00:26:57.940
you've trained your tree, you get some
00:26:57.940 --> 00:26:59.450
new test sample and you want to know is
00:26:59.450 --> 00:27:01.450
that a lemon or an orange kind of looks
00:27:01.450 --> 00:27:01.920
in between.
00:27:02.610 --> 00:27:05.295
So you is it greater than 6.5
00:27:05.295 --> 00:27:05.740
centimeters?
00:27:05.740 --> 00:27:06.185
No.
00:27:06.185 --> 00:27:08.355
Is a tight greater than 6 centimeters?
00:27:08.355 --> 00:27:08.690
No.
00:27:08.690 --> 00:27:10.110
And so therefore it's an orange
00:27:10.110 --> 00:27:10.970
according to your rule.
00:27:13.260 --> 00:27:15.053
And you could take this tree and could
00:27:15.053 --> 00:27:17.456
you could rewrite it as a set of rules,
00:27:17.456 --> 00:27:20.560
like one rule is greater than 6.5,
00:27:20.560 --> 00:27:23.478
height greater than 9.5, another rule
00:27:23.478 --> 00:27:26.020
is greater than 65, height less than
00:27:26.020 --> 00:27:27.330
9.5, and so on.
00:27:27.330 --> 00:27:28.640
There's like 4 different rules
00:27:28.640 --> 00:27:31.180
represented by this tree, and each rule
00:27:31.180 --> 00:27:33.950
corresponds to some section of the
00:27:33.950 --> 00:27:36.440
feature space, and each rule yields
00:27:36.440 --> 00:27:37.140
some prediction.
00:27:40.950 --> 00:27:44.020
So here's another example with some
00:27:44.020 --> 00:27:45.580
discrete inputs.
00:27:45.580 --> 00:27:48.030
So here the prediction problem is to
00:27:48.030 --> 00:27:49.955
tell whether or not somebody's going to
00:27:49.955 --> 00:27:50.350
wait.
00:27:50.350 --> 00:27:52.440
If they go to a restaurant and they're
00:27:52.440 --> 00:27:54.173
told they have to wait, so do they wait
00:27:54.173 --> 00:27:55.160
or do they leave?
00:27:56.290 --> 00:27:58.170
And the features are things like
00:27:58.170 --> 00:28:00.160
whether there's an alternative nearby,
00:28:00.160 --> 00:28:02.240
whether there's a bar they can wait at,
00:28:02.240 --> 00:28:03.900
whether it's Friday or Saturday,
00:28:03.900 --> 00:28:05.289
whether they're Hungry, whether the
00:28:05.290 --> 00:28:07.106
restaurants full, what the price is,
00:28:07.106 --> 00:28:08.740
whether it's raining, whether they had
00:28:08.740 --> 00:28:10.560
a Reservation, what type of restaurant
00:28:10.560 --> 00:28:12.900
is, and they would wait time.
00:28:12.900 --> 00:28:14.747
And these are all categorical, so the
00:28:14.747 --> 00:28:16.100
wait time is split into different
00:28:16.100 --> 00:28:16.540
chunks.
00:28:20.660 --> 00:28:22.670
And so you could.
00:28:24.110 --> 00:28:27.810
You could train a tree from these
00:28:27.810 --> 00:28:29.820
categorical variables, and of course I
00:28:29.820 --> 00:28:31.590
will tell you more about like how you
00:28:31.590 --> 00:28:32.390
would learn this tree.
00:28:33.960 --> 00:28:35.670
But you might have a tree like this
00:28:35.670 --> 00:28:36.500
where you say.
00:28:37.730 --> 00:28:39.770
First, are there are there people in
00:28:39.770 --> 00:28:40.370
the restaurant?
00:28:40.370 --> 00:28:41.800
Patrons means like it's a restaurant
00:28:41.800 --> 00:28:42.684
full or not.
00:28:42.684 --> 00:28:46.310
If it's not full, then you leave right
00:28:46.310 --> 00:28:47.790
away because they're just being rude.
00:28:47.790 --> 00:28:49.330
If they tell, you have to wait I guess.
00:28:49.930 --> 00:28:52.140
If it's partly full then you'll wait,
00:28:52.140 --> 00:28:54.080
and if it's full then you then you have
00:28:54.080 --> 00:28:55.680
like consider further things.
00:28:55.680 --> 00:28:58.360
If it's a WaitEstimate, really short,
00:28:58.360 --> 00:28:58.990
then you wait.
00:28:58.990 --> 00:28:59.660
Is it really long?
00:28:59.660 --> 00:29:00.170
Then you don't.
00:29:00.960 --> 00:29:03.290
Otherwise, are you hungry?
00:29:03.290 --> 00:29:04.693
If you're not, then you'll wait.
00:29:04.693 --> 00:29:06.600
If you are, then you keep thinking.
00:29:06.600 --> 00:29:08.320
So you have like, all this series of
00:29:08.320 --> 00:29:08.820
choices.
00:29:10.350 --> 00:29:12.790
That these trees and practice like if
00:29:12.790 --> 00:29:14.230
you were to use a Decision tree on
00:29:14.230 --> 00:29:14.680
MNIST.
00:29:15.810 --> 00:29:17.600
Where the features are pretty weak
00:29:17.600 --> 00:29:19.510
individually, they're just like pixel
00:29:19.510 --> 00:29:20.140
values.
00:29:20.140 --> 00:29:21.610
You can imagine that this tree could
00:29:21.610 --> 00:29:23.160
get really complicated and long.
00:29:27.970 --> 00:29:28.390
Right.
00:29:28.390 --> 00:29:31.840
So just to mostly be state.
00:29:32.450 --> 00:29:34.080
And then Decision tree.
00:29:34.080 --> 00:29:36.410
The internal nodes are Test Attributes,
00:29:36.410 --> 00:29:38.150
so it's some kind of like feature.
00:29:38.150 --> 00:29:40.050
Attribute and feature are synonymous,
00:29:40.050 --> 00:29:41.110
they're the same thing.
00:29:41.830 --> 00:29:45.880
Some kind of feature attribute and.
00:29:45.960 --> 00:29:47.420
And if it's a continuous attribute then
00:29:47.420 --> 00:29:48.420
you have to have some kind of
00:29:48.420 --> 00:29:53.420
threshold, so width greater than 6.5 or
00:29:53.420 --> 00:29:54.650
is it raining or not?
00:29:54.650 --> 00:29:56.050
Those are two examples of.
00:29:56.740 --> 00:29:57.440
Of tests.
00:29:58.370 --> 00:29:59.984
Then depending on the outcome of that
00:29:59.984 --> 00:30:02.310
test, you split in different ways, and
00:30:02.310 --> 00:30:03.860
when you're Training, you split all
00:30:03.860 --> 00:30:05.935
your data according to that test, and
00:30:05.935 --> 00:30:07.960
then you're going to solve again within
00:30:07.960 --> 00:30:09.390
each of those nodes separately.
00:30:10.480 --> 00:30:11.570
For the next Test.
00:30:12.260 --> 00:30:14.110
Until you get to a leaf node, and at
00:30:14.110 --> 00:30:16.125
the leaf node you provide an output or
00:30:16.125 --> 00:30:18.532
a prediction, which could be, which in
00:30:18.532 --> 00:30:20.480
this case is a class, in this
00:30:20.480 --> 00:30:21.780
particular example whether it's a
00:30:21.780 --> 00:30:22.540
Linear orange.
00:30:25.060 --> 00:30:25.260
Yep.
00:30:29.360 --> 00:30:31.850
So the question is how does it Decision
00:30:31.850 --> 00:30:34.700
tree account for anomalies as in late
00:30:34.700 --> 00:30:36.480
mislabeled data or really weird
00:30:36.480 --> 00:30:37.260
examples or?
00:30:50.100 --> 00:30:52.370
So the so the question is like how does
00:30:52.370 --> 00:30:54.400
it Decision tree deal with weird or
00:30:54.400 --> 00:30:55.860
unlikely examples?
00:30:55.860 --> 00:30:58.020
And that's a good question because one
00:30:58.020 --> 00:31:00.200
of the things about a Decision tree is
00:31:00.200 --> 00:31:01.510
that if you train it.
00:31:02.350 --> 00:31:04.460
If you train it, if you train the full
00:31:04.460 --> 00:31:06.560
tree, then you can always.
00:31:06.560 --> 00:31:09.970
As long as the feature vectors for each
00:31:09.970 --> 00:31:11.560
sample are unique, you can always get
00:31:11.560 --> 00:31:13.470
perfect Classification Error.
00:31:13.470 --> 00:31:14.900
A tree has no bias.
00:31:14.900 --> 00:31:16.580
You can always like fit your training
00:31:16.580 --> 00:31:18.980
data perfectly because you just keep on
00:31:18.980 --> 00:31:20.530
chopping it into smaller and smaller
00:31:20.530 --> 00:31:22.070
bits until finally the answer.
00:31:22.800 --> 00:31:24.960
So as a result, that can be dangerous
00:31:24.960 --> 00:31:26.767
because if you do have some unusual
00:31:26.767 --> 00:31:29.100
examples, you can end up creating rules
00:31:29.100 --> 00:31:31.410
based on those examples that don't
00:31:31.410 --> 00:31:32.920
generalize well tuning data.
00:31:33.640 --> 00:31:36.191
And so some things that you can do are
00:31:36.191 --> 00:31:38.119
you can stop Training, stop Training
00:31:38.120 --> 00:31:38.590
early.
00:31:38.590 --> 00:31:40.440
So you can say I'm not going to split
00:31:40.440 --> 00:31:42.530
once I only have 5 examples of my leaf
00:31:42.530 --> 00:31:43.990
node, I'm going to quit splitting and
00:31:43.990 --> 00:31:45.460
I'll just output my best guess.
00:31:46.520 --> 00:31:47.070
00:31:47.990 --> 00:31:49.240
There's also like.
00:31:52.250 --> 00:31:53.770
Probably on Tuesday.
00:31:53.770 --> 00:31:54.810
Actually, I'm going to talk about
00:31:54.810 --> 00:31:56.790
ensembles, which is ways of combining
00:31:56.790 --> 00:31:58.770
money trees, which is another way of
00:31:58.770 --> 00:31:59.840
getting rid of this problem.
00:32:01.360 --> 00:32:01.750
Question.
00:32:09.850 --> 00:32:11.190
That's a good question too.
00:32:11.190 --> 00:32:12.890
So the question is whether Decision
00:32:12.890 --> 00:32:14.500
trees are always binary.
00:32:14.500 --> 00:32:17.880
So like in this example, it's not
00:32:17.880 --> 00:32:21.640
binary, they're splitting like the
00:32:21.640 --> 00:32:23.250
Patrons is splitting based on three
00:32:23.250 --> 00:32:23.780
values.
00:32:24.510 --> 00:32:28.260
But typically they are binary.
00:32:28.260 --> 00:32:29.030
So if you're.
00:32:29.810 --> 00:32:32.277
If you're using continuous values, it
00:32:32.277 --> 00:32:32.535
will.
00:32:32.535 --> 00:32:34.410
It will almost always be binary,
00:32:34.410 --> 00:32:35.440
because you could.
00:32:35.440 --> 00:32:37.330
Even if you wanted to split continuous
00:32:37.330 --> 00:32:40.260
variables into many different chunks,
00:32:40.260 --> 00:32:42.350
you can do that through a sequence of
00:32:42.350 --> 00:32:43.380
binary decisions.
00:32:44.780 --> 00:32:47.620
In SK learn as well, their Decision
00:32:47.620 --> 00:32:50.160
trees cannot deal with like multi
00:32:50.160 --> 00:32:53.040
valued attributes and so you need to
00:32:53.040 --> 00:32:55.000
convert them into binary attributes in
00:32:55.000 --> 00:32:56.470
order to use sklearn.
00:32:57.350 --> 00:32:59.470
And I think often that's done as a
00:32:59.470 --> 00:33:01.590
design Decision, because otherwise like
00:33:01.590 --> 00:33:03.050
some features will be like
00:33:03.050 --> 00:33:04.780
intrinsically more powerful than other
00:33:04.780 --> 00:33:06.690
features if they create like more
00:33:06.690 --> 00:33:07.340
splits.
00:33:07.340 --> 00:33:09.160
So it can cause like a bias in your
00:33:09.160 --> 00:33:09.990
feature selection.
00:33:10.740 --> 00:33:12.990
So they don't have to be binary, but
00:33:12.990 --> 00:33:15.160
it's a common common setting.
00:33:21.880 --> 00:33:24.480
Alright, so the Training 4 Decision
00:33:24.480 --> 00:33:26.760
tree again without yet getting into the
00:33:26.760 --> 00:33:27.000
math.
00:33:27.710 --> 00:33:30.935
Is Recursively for each node in the
00:33:30.935 --> 00:33:32.590
tree, if the labels and the node are
00:33:32.590 --> 00:33:33.120
mixed.
00:33:33.120 --> 00:33:35.256
So to start with, we're at the of the
00:33:35.256 --> 00:33:38.030
tree and we have all this data, and so
00:33:38.030 --> 00:33:39.676
essentially there's just right now some
00:33:39.676 --> 00:33:41.025
probability that's a no, some
00:33:41.025 --> 00:33:41.950
probability that's an 784x1.
00:33:41.950 --> 00:33:43.900
Those probabilities are close to 5050.
00:33:45.530 --> 00:33:47.410
Then I'm going to choose some attribute
00:33:47.410 --> 00:33:50.020
and split the values based on the data
00:33:50.020 --> 00:33:51.200
that reaches that node.
00:33:52.310 --> 00:33:54.310
So here I Choose this attribute the
00:33:54.310 --> 00:33:55.430
tree I'm creating up there.
00:33:56.110 --> 00:33:58.060
X2 is less than .6.
00:34:00.630 --> 00:34:05.310
If it's less than .6 then I go down one
00:34:05.310 --> 00:34:07.067
branch and if it's greater than I go
00:34:07.067 --> 00:34:08.210
down the other branch.
00:34:08.210 --> 00:34:10.870
So now then I can now start making
00:34:10.870 --> 00:34:13.440
decisions separately about this region
00:34:13.440 --> 00:34:14.260
in this region.
00:34:15.910 --> 00:34:19.200
So then I Choose another node and I say
00:34:19.200 --> 00:34:21.660
if X1 is less than 7.
00:34:22.630 --> 00:34:24.360
So I create this split and this only
00:34:24.360 --> 00:34:25.490
pertains to the data.
00:34:25.490 --> 00:34:27.570
Now that came down the first node so
00:34:27.570 --> 00:34:28.989
it's this side of the data.
00:34:29.710 --> 00:34:31.292
So if it's over here, then it's a no,
00:34:31.292 --> 00:34:33.220
if it's over here, then it's an X and
00:34:33.220 --> 00:34:34.620
Now I don't need to create anymore
00:34:34.620 --> 00:34:36.690
Decision nodes for this whole region of
00:34:36.690 --> 00:34:38.870
space because I have perfect
00:34:38.870 --> 00:34:39.630
Classification.
00:34:40.760 --> 00:34:43.010
Then I go to my top side.
00:34:43.730 --> 00:34:45.390
And I can make another split.
00:34:45.390 --> 00:34:47.760
So here there's actually more than one
00:34:47.760 --> 00:34:48.015
choice.
00:34:48.015 --> 00:34:49.960
I think that's like kind of equally
00:34:49.960 --> 00:34:51.460
good, but.
00:34:51.570 --> 00:34:56.230
Again, say if X2 is less than .8, then
00:34:56.230 --> 00:34:57.960
it goes down here where I'm still
00:34:57.960 --> 00:34:58.250
unsure.
00:34:58.250 --> 00:35:00.080
If it's greater than eight, then it's
00:35:00.080 --> 00:35:00.980
definitely a red X.
00:35:03.260 --> 00:35:05.120
And then I can keep doing that until I
00:35:05.120 --> 00:35:07.190
finally have a perfect Classification
00:35:07.190 --> 00:35:08.030
in the training data.
00:35:08.810 --> 00:35:10.000
So that's the full tree.
00:35:11.070 --> 00:35:13.830
And if you could stop early, you could
00:35:13.830 --> 00:35:15.739
say I'm not going to go past like 3
00:35:15.740 --> 00:35:18.310
levels, or that I'm going to stop
00:35:18.310 --> 00:35:20.910
splitting once my leaf node doesn't
00:35:20.910 --> 00:35:23.210
have more than five examples.
00:35:39.470 --> 00:35:41.560
Well, the question was does the first
00:35:41.560 --> 00:35:42.320
split matter?
00:35:42.320 --> 00:35:43.929
So I guess there's two parts to that.
00:35:43.930 --> 00:35:45.880
One is that I will tell you how we do
00:35:45.880 --> 00:35:46.980
this computationally.
00:35:46.980 --> 00:35:48.780
So you try to greedily find like the
00:35:48.780 --> 00:35:49.900
best split every time.
00:35:50.990 --> 00:35:53.530
And the other thing is that finding the
00:35:53.530 --> 00:35:57.290
minimum size tree is like a
00:35:57.290 --> 00:35:59.540
computationally hard problem.
00:36:00.540 --> 00:36:01.766
So it's infeasible.
00:36:01.766 --> 00:36:04.530
So you end up with a greedy solution
00:36:04.530 --> 00:36:06.020
where for every node you're choosing
00:36:06.020 --> 00:36:08.200
the best split for that node.
00:36:08.200 --> 00:36:10.045
But that doesn't necessarily give you
00:36:10.045 --> 00:36:11.680
the shortest tree overall, because you
00:36:11.680 --> 00:36:13.020
don't know like what kinds of splits
00:36:13.020 --> 00:36:14.250
will be available to you later.
00:36:16.710 --> 00:36:19.050
So it does matter, but you have like
00:36:19.050 --> 00:36:20.630
there's an algorithm for doing it in a
00:36:20.630 --> 00:36:21.550
decent way, yeah.
00:36:55.320 --> 00:36:55.860
00:36:57.660 --> 00:36:59.080
There have well.
00:37:01.160 --> 00:37:02.650
How will you know that it will work for
00:37:02.650 --> 00:37:03.320
like new data?
00:37:05.000 --> 00:37:09.209
So basically if you want to know, you
00:37:09.210 --> 00:37:10.740
do always want to know, you always want
00:37:10.740 --> 00:37:12.420
to know, right, if you if the model
00:37:12.420 --> 00:37:13.620
that you learned is going to work for
00:37:13.620 --> 00:37:14.540
new data.
00:37:14.540 --> 00:37:16.370
And so that's why I typically you would
00:37:16.370 --> 00:37:18.030
carve off, if you have some Training
00:37:18.030 --> 00:37:19.800
set, you'd carve off a validation set.
00:37:20.450 --> 00:37:22.380
And you would train it say with like
00:37:22.380 --> 00:37:25.040
70% of the Training examples and test
00:37:25.040 --> 00:37:27.850
it on the 30% of the held out Samples?
00:37:28.530 --> 00:37:30.040
And then there was held out Samples
00:37:30.040 --> 00:37:32.170
will give you an estimate of how well
00:37:32.170 --> 00:37:33.260
your method works.
00:37:33.260 --> 00:37:35.250
And so then like if you find for
00:37:35.250 --> 00:37:37.626
example that I trained a full tree and
00:37:37.626 --> 00:37:39.850
of course I got like 0% Training error,
00:37:39.850 --> 00:37:41.850
but my Test error is like 40%.
00:37:42.590 --> 00:37:44.990
Then you would probably say maybe I
00:37:44.990 --> 00:37:46.803
should try Training a shorter tree and
00:37:46.803 --> 00:37:48.930
then you can like retrain it with some
00:37:48.930 --> 00:37:51.120
constraints and then test it again on
00:37:51.120 --> 00:37:52.755
your validation set and Choose like
00:37:52.755 --> 00:37:53.830
your Parameters that way.
00:37:54.860 --> 00:37:57.581
There's also I'll talk about most
00:37:57.581 --> 00:37:59.140
likely, most likely this.
00:37:59.140 --> 00:38:01.020
I was planning to do it Thursday, but
00:38:01.020 --> 00:38:02.187
I'll probably do it next Tuesday.
00:38:02.187 --> 00:38:04.580
I'll talk about ensembles, including
00:38:04.580 --> 00:38:06.622
random forests, and those are like kind
00:38:06.622 --> 00:38:09.150
of like brain dead always work methods
00:38:09.150 --> 00:38:11.080
that combine a lot of trees and are
00:38:11.080 --> 00:38:13.439
really reliable whether you have a lot
00:38:13.439 --> 00:38:16.087
of data or well, you kind of need data.
00:38:16.087 --> 00:38:17.665
But whether you have a lot of features
00:38:17.665 --> 00:38:19.850
or a little features, they always work.
00:38:20.480 --> 00:38:21.480
They always work pretty well.
00:38:23.820 --> 00:38:25.950
Right, so in prediction then you just
00:38:25.950 --> 00:38:27.560
basically descend the tree, so you
00:38:27.560 --> 00:38:29.920
check the conditions is tX2 greater
00:38:29.920 --> 00:38:32.370
than .6 blah blah blah blah blah until
00:38:32.370 --> 00:38:33.750
you find yourself in a leaf node.
00:38:34.380 --> 00:38:36.630
So for example, if I have this data
00:38:36.630 --> 00:38:38.500
point and I'm trying to classify it, I
00:38:38.500 --> 00:38:40.902
would end up following these rules down
00:38:40.902 --> 00:38:44.290
to down to the leaf node of.
00:38:45.960 --> 00:38:47.418
Yeah, like right over here, right?
00:38:47.418 --> 00:38:50.158
X2 is less than .6 and X1 is less than
00:38:50.158 --> 00:38:50.460
.7.
00:38:51.260 --> 00:38:52.740
And so that's going to be no.
00:38:53.860 --> 00:38:56.500
And if I am over here then I end up
00:38:56.500 --> 00:38:59.420
following going down to here to here.
00:39:00.480 --> 00:39:03.020
To here to here and I end up in this
00:39:03.020 --> 00:39:07.299
leaf node and so it's an X and it
00:39:07.300 --> 00:39:09.395
doesn't matter like where it falls in
00:39:09.395 --> 00:39:10.580
this part of the space.
00:39:10.580 --> 00:39:11.700
Usually this isn't like.
00:39:12.390 --> 00:39:13.770
Even something you necessarily
00:39:13.770 --> 00:39:15.520
visualize, but.
00:39:16.060 --> 00:39:18.025
But it's worth noting that even parts
00:39:18.025 --> 00:39:20.020
of your feature space that are kind of
00:39:20.020 --> 00:39:22.640
far away from any Example can still get
00:39:22.640 --> 00:39:24.360
classified by this Decision tree.
00:39:25.070 --> 00:39:27.670
And it's not necessarily the Nearest
00:39:27.670 --> 00:39:28.450
neighbor Decision.
00:39:28.450 --> 00:39:31.186
Like this star here is actually closer
00:39:31.186 --> 00:39:33.390
to the 784x1 than it is to the O's, but
00:39:33.390 --> 00:39:35.010
it would still be a no because it's on
00:39:35.010 --> 00:39:36.050
that side of the boundary.
00:39:40.650 --> 00:39:42.350
So the key question is, how do you
00:39:42.350 --> 00:39:45.810
choose what attribute to split and
00:39:45.810 --> 00:39:46.384
where to split?
00:39:46.384 --> 00:39:48.390
So how do you decide what test you're
00:39:48.390 --> 00:39:50.000
going to use for a given node?
00:39:50.920 --> 00:39:53.615
And so let's take this example.
00:39:53.615 --> 00:39:56.290
So here I've got some table of features
00:39:56.290 --> 00:39:57.180
and predictions.
00:39:58.020 --> 00:39:59.010
And if.
00:40:00.410 --> 00:40:02.280
And if I were to split, these are
00:40:02.280 --> 00:40:04.570
binary features so they just have two
00:40:04.570 --> 00:40:06.430
values T2 false I guess.
00:40:07.400 --> 00:40:07.940
If.
00:40:09.620 --> 00:40:12.440
If I split based on X1 and I go in One
00:40:12.440 --> 00:40:14.570
Direction, then it's all true.
00:40:15.200 --> 00:40:17.585
The prediction is true and if I go in
00:40:17.585 --> 00:40:19.920
the other direction then 3/4 of the
00:40:19.920 --> 00:40:21.080
time the prediction is false.
00:40:22.810 --> 00:40:26.343
If I split based on X2, then 3/4 of the
00:40:26.343 --> 00:40:27.948
time the prediction is true.
00:40:27.948 --> 00:40:31.096
If it's true and 50% of the time the
00:40:31.096 --> 00:40:32.819
prediction is false, X2 is false.
00:40:33.530 --> 00:40:36.300
So which of these features is a better
00:40:36.300 --> 00:40:37.400
Test?
00:40:39.550 --> 00:40:41.530
So how many people think that the left
00:40:41.530 --> 00:40:42.530
is a better Test?
00:40:43.790 --> 00:40:45.070
How many people think they're right is
00:40:45.070 --> 00:40:45.730
a better Test.
00:40:46.840 --> 00:40:48.380
Right the left is a better Test
00:40:48.380 --> 00:40:48.950
because.
00:40:50.620 --> 00:40:53.380
Because my uncertainty is greatly
00:40:53.380 --> 00:40:54.990
reduced on the left side.
00:40:54.990 --> 00:40:58.750
So initially, initially I had like a
00:40:58.750 --> 00:41:01.280
5/8 chance of getting it right if I
00:41:01.280 --> 00:41:02.280
just guessed true.
00:41:02.910 --> 00:41:06.706
But if I know X1, then I've got a 100%
00:41:06.706 --> 00:41:08.600
chance of getting it right, at least in
00:41:08.600 --> 00:41:09.494
the training data.
00:41:09.494 --> 00:41:13.338
If I know that X1 is true, and I've got
00:41:13.338 --> 00:41:15.132
a 3/4 chance of getting it right if I
00:41:15.132 --> 00:41:16.280
know that X1 is false.
00:41:16.280 --> 00:41:19.135
So X1 tells me a lot about the
00:41:19.135 --> 00:41:19.572
prediction.
00:41:19.572 --> 00:41:22.035
It greatly reduces my uncertainty about
00:41:22.035 --> 00:41:22.890
the prediction.
00:41:24.510 --> 00:41:26.412
And to quantify this, we need to
00:41:26.412 --> 00:41:28.560
quantify uncertainty and then be able
00:41:28.560 --> 00:41:32.350
to measure how much a certain feature
00:41:32.350 --> 00:41:33.950
reduces our uncertainty in the
00:41:33.950 --> 00:41:34.720
prediction.
00:41:34.720 --> 00:41:36.800
And that's called the information gain.
00:41:40.470 --> 00:41:44.540
So to quantify the uncertainty, I'll
00:41:44.540 --> 00:41:45.595
use these two examples.
00:41:45.595 --> 00:41:47.790
So imagine that you're flipping a coin.
00:41:47.790 --> 00:41:50.150
These are like heads and tails, or
00:41:50.150 --> 00:41:51.510
present them as zeros and ones.
00:41:52.180 --> 00:41:54.820
And so one time I've got two different
00:41:54.820 --> 00:41:56.186
sequences, let's say two different
00:41:56.186 --> 00:41:57.740
coins and one in the coins.
00:41:57.740 --> 00:42:00.120
It's a biased coin, so I end up with
00:42:00.120 --> 00:42:03.330
zeros or heads like 16 out of 18 times.
00:42:04.250 --> 00:42:06.520
And the other for the other Coin I get
00:42:06.520 --> 00:42:09.400
closer to 5058 out of.
00:42:10.050 --> 00:42:12.390
18 times I get heads so.
00:42:13.530 --> 00:42:17.520
Which of these has higher uncertainty?
00:42:18.540 --> 00:42:19.730
The left or the right?
00:42:21.330 --> 00:42:22.580
Right, correct.
00:42:22.580 --> 00:42:23.070
They're right.
00:42:23.070 --> 00:42:24.900
Has a lot higher uncertainty.
00:42:24.900 --> 00:42:27.370
So if I with that Coin, I really don't
00:42:27.370 --> 00:42:28.470
know if it's going to be heads or
00:42:28.470 --> 00:42:30.860
tails, but on the left side, I'm pretty
00:42:30.860 --> 00:42:31.820
sure it's going to be heads.
00:42:32.590 --> 00:42:33.360
Or zeros.
00:42:34.720 --> 00:42:36.770
So we can measure that with this
00:42:36.770 --> 00:42:38.645
function called Entropy.
00:42:38.645 --> 00:42:41.350
So the entropy is a measure of
00:42:41.350 --> 00:42:42.030
uncertainty.
00:42:42.960 --> 00:42:45.740
And it's defined as the negative sum
00:42:45.740 --> 00:42:48.070
over all the values of some variable of
00:42:48.070 --> 00:42:50.220
the probability of that value.
00:42:51.020 --> 00:42:53.490
Times the log probability that value,
00:42:53.490 --> 00:42:56.470
and people usually sometimes use like
00:42:56.470 --> 00:42:57.520
log base 2.
00:42:58.630 --> 00:43:00.700
Just because that way the Entropy
00:43:00.700 --> 00:43:02.550
ranges from zero to 1 if you have
00:43:02.550 --> 00:43:03.490
binary variables.
00:43:07.600 --> 00:43:10.820
So for this case here, the Entropy
00:43:10.820 --> 00:43:13.280
would be -, 8 ninths, because eight out
00:43:13.280 --> 00:43:14.600
of nine times it's zero.
00:43:15.270 --> 00:43:17.300
Times log two of eight ninths.
00:43:18.230 --> 00:43:21.210
Minus one ninth times, log 2 of 1 ninth
00:43:21.210 --> 00:43:22.790
and that works out to about 1/2.
00:43:24.370 --> 00:43:28.480
And over here the Entropy is -, 4
00:43:28.480 --> 00:43:30.270
ninths because four out of nine times,
00:43:30.270 --> 00:43:32.900
or 8 out of 18 times, it's a 0.
00:43:34.410 --> 00:43:37.104
Times log 24 ninths, minus five ninths,
00:43:37.104 --> 00:43:39.010
times log two of five ninths, and
00:43:39.010 --> 00:43:41.490
that's about 99.
00:43:43.430 --> 00:43:45.280
The Entropy measure is how surprised
00:43:45.280 --> 00:43:47.595
are we by some new value of this
00:43:47.595 --> 00:43:47.830
Sequence?
00:43:47.830 --> 00:43:50.123
How surprised are we likely to be in,
00:43:50.123 --> 00:43:52.460
or how much information does it convey
00:43:52.460 --> 00:43:54.895
that we know that we're in this
00:43:54.895 --> 00:43:56.974
Sequence, or more generally, that we
00:43:56.974 --> 00:43:57.940
know some feature?
00:44:01.100 --> 00:44:03.425
So this is just showing the Entropy if
00:44:03.425 --> 00:44:05.450
the probability if you have a binary
00:44:05.450 --> 00:44:06.340
variable X.
00:44:07.110 --> 00:44:09.720
And the probability of X is 0, then
00:44:09.720 --> 00:44:12.180
your Entropy is 0 because you always
00:44:12.180 --> 00:44:15.127
know that if probability of X = 2 is
00:44:15.127 --> 00:44:16.818
zero, that means that probability of X
00:44:16.818 --> 00:44:18.210
equals false is 1.
00:44:18.860 --> 00:44:20.530
And so therefore you have complete
00:44:20.530 --> 00:44:22.470
confidence that the value will be
00:44:22.470 --> 00:44:22.810
false.
00:44:24.070 --> 00:44:27.740
If probability of X is true is 1, then
00:44:27.740 --> 00:44:29.590
you have complete confidence that the
00:44:29.590 --> 00:44:30.650
value will be true.
00:44:31.440 --> 00:44:35.570
But if it's .5, then you have no
00:44:35.570 --> 00:44:37.120
information about whether it's true or
00:44:37.120 --> 00:44:39.520
false, and so you have maximum entropy,
00:44:39.520 --> 00:44:40.190
which is 1.
00:44:45.770 --> 00:44:47.280
So here's another example.
00:44:47.280 --> 00:44:49.340
So suppose that we've got two
00:44:49.340 --> 00:44:51.070
variables, whether it's raining or not,
00:44:51.070 --> 00:44:52.220
and whether it's cloudy or not.
00:44:52.820 --> 00:44:55.700
And we've observed 100 days and marked
00:44:55.700 --> 00:44:57.260
down whether it's rainy or cloudy.
00:44:58.870 --> 00:45:00.150
Many and or Cloudy.
00:45:00.930 --> 00:45:01.500
00:45:02.600 --> 00:45:06.300
So 24 days it was raining and cloudy.
00:45:06.300 --> 00:45:08.210
One day it was raining and not Cloudy.
00:45:09.320 --> 00:45:11.244
25 days it was not raining and cloudy
00:45:11.244 --> 00:45:13.409
and 50 days it was not raining and not
00:45:13.409 --> 00:45:13.649
Cloudy.
00:45:15.620 --> 00:45:17.980
The probabilities are just dividing by
00:45:17.980 --> 00:45:18.766
the total there.
00:45:18.766 --> 00:45:20.850
So the probability of Cloudy and not
00:45:20.850 --> 00:45:22.630
raining is 25 out of 100.
00:45:24.040 --> 00:45:26.660
And so I can also compute an Entropy of
00:45:26.660 --> 00:45:27.855
this whole joint distribution.
00:45:27.855 --> 00:45:31.150
So I can say that the entropy of X&Y
00:45:31.150 --> 00:45:33.446
together is the sum all the different
00:45:33.446 --> 00:45:35.428
values of X and the over all the
00:45:35.428 --> 00:45:36.419
different values of Y.
00:45:37.060 --> 00:45:39.770
Of probability of X&Y times log 2,
00:45:39.770 --> 00:45:41.920
probability of X&Y, and then that's all
00:45:41.920 --> 00:45:42.880
just like written out here.
00:45:43.650 --> 00:45:45.115
And then I get some Entropy value.
00:45:45.115 --> 00:45:47.940
And sometimes people call those units
00:45:47.940 --> 00:45:51.490
bits, so 156 bits because that's the
00:45:51.490 --> 00:45:53.008
amount of, that's the number of bits
00:45:53.008 --> 00:45:54.680
that I would need that I would expect
00:45:54.680 --> 00:45:55.040
to.
00:45:55.790 --> 00:45:57.780
Be able to like represent this.
00:45:58.630 --> 00:45:59.700
This information.
00:46:00.430 --> 00:46:04.395
If you if it were always not Cloudy and
00:46:04.395 --> 00:46:04.990
not raining.
00:46:05.850 --> 00:46:08.020
If it were 100% of the time not Cloudy
00:46:08.020 --> 00:46:10.280
and not raining, then you'd have 0 bits
00:46:10.280 --> 00:46:11.830
because you don't need any data to
00:46:11.830 --> 00:46:12.810
represent the.
00:46:13.710 --> 00:46:15.770
That uncertainty, it's just always
00:46:15.770 --> 00:46:16.300
true.
00:46:16.300 --> 00:46:18.300
I mean it's always like one value.
00:46:18.300 --> 00:46:20.790
So 15 bits means that you have pretty
00:46:20.790 --> 00:46:21.490
high uncertainty.
00:46:25.250 --> 00:46:27.680
There's also a concept called specific
00:46:27.680 --> 00:46:28.510
Entropy.
00:46:28.510 --> 00:46:29.780
So that is.
00:46:29.780 --> 00:46:33.560
That means that if one thing, then how
00:46:33.560 --> 00:46:34.516
much does that?
00:46:34.516 --> 00:46:36.610
How much uncertainty do you have left?
00:46:37.460 --> 00:46:41.170
So, for example, what is the entropy of
00:46:41.170 --> 00:46:43.610
cloudiness given that I know that it's
00:46:43.610 --> 00:46:44.000
raining?
00:46:45.420 --> 00:46:48.940
And the Conditional Entropy is very
00:46:48.940 --> 00:46:51.280
similar form, it's just negative sum
00:46:51.280 --> 00:46:52.720
over the values of the.
00:46:53.710 --> 00:46:54.970
The thing that you're measuring the
00:46:54.970 --> 00:46:55.780
Entropy over.
00:46:56.800 --> 00:46:58.880
The probability of that given the thing
00:46:58.880 --> 00:46:59.500
that.
00:47:00.150 --> 00:47:03.610
Times the log probability of Y given X,
00:47:03.610 --> 00:47:04.760
where Y is the thing you're measuring
00:47:04.760 --> 00:47:06.400
the uncertainty of, and X is a thing
00:47:06.400 --> 00:47:06.850
that you know.
00:47:09.200 --> 00:47:12.660
So if I know that it's Cloudy, then
00:47:12.660 --> 00:47:15.690
there's a 24 out of 25 chance that
00:47:15.690 --> 00:47:16.150
it's.
00:47:17.340 --> 00:47:17.950
Wait, no.
00:47:17.950 --> 00:47:19.910
If I know that it's raining, sorry.
00:47:19.910 --> 00:47:21.599
If I know that it's raining, then
00:47:21.600 --> 00:47:23.931
there's a 24 out of 25 chance that it's
00:47:23.931 --> 00:47:24.430
Cloudy, right?
00:47:24.430 --> 00:47:26.190
And then one out of 25 chance that it's
00:47:26.190 --> 00:47:26.760
not Cloudy.
00:47:27.600 --> 00:47:30.340
So I get 24 to 25 there and one out of
00:47:30.340 --> 00:47:33.070
25 there, and now my Entropy is greatly
00:47:33.070 --> 00:47:33.660
reduced.
00:47:39.810 --> 00:47:41.280
And then you can also measure.
00:47:41.930 --> 00:47:44.250
In expected Conditional Entropy.
00:47:46.020 --> 00:47:50.570
So that's just the probability of.
00:47:50.570 --> 00:47:53.870
That's just taking the specific
00:47:53.870 --> 00:47:54.940
Conditional Entropy.
00:47:55.780 --> 00:47:58.660
At times the probability of each of the
00:47:58.660 --> 00:48:00.360
values that I might know.
00:48:01.260 --> 00:48:03.710
Summed up over the different values,
00:48:03.710 --> 00:48:04.180
so.
00:48:04.900 --> 00:48:06.130
The.
00:48:06.820 --> 00:48:09.920
The expected Conditional value Entropy
00:48:09.920 --> 00:48:11.950
for knowing whether or not it's raining
00:48:11.950 --> 00:48:15.179
would be the Conditional Entropy.
00:48:16.040 --> 00:48:19.010
Of it raining if I know it's raining.
00:48:19.920 --> 00:48:21.280
Times the probability that it's
00:48:21.280 --> 00:48:21.720
raining.
00:48:22.460 --> 00:48:24.460
Plus the.
00:48:25.190 --> 00:48:28.270
Entropy of cloudiness given that it's
00:48:28.270 --> 00:48:30.210
not raining, times the probability
00:48:30.210 --> 00:48:30.900
that's not raining.
00:48:33.530 --> 00:48:35.550
And that's also equal to this thing.
00:48:42.960 --> 00:48:43.400
Right.
00:48:43.400 --> 00:48:46.168
So if I want to know what is the
00:48:46.168 --> 00:48:47.790
entropy of cloudiness, I guess I said
00:48:47.790 --> 00:48:48.730
it a little early.
00:48:48.730 --> 00:48:50.890
What is the entropy of cloudiness given
00:48:50.890 --> 00:48:52.720
whether that we know whether or not
00:48:52.720 --> 00:48:53.340
it's raining?
00:48:54.310 --> 00:48:56.240
Then that is.
00:48:56.850 --> 00:48:59.540
Going to be like 1/4, which is the
00:48:59.540 --> 00:49:02.009
probability that it's raining, is that
00:49:02.010 --> 00:49:02.320
right?
00:49:02.320 --> 00:49:04.790
25 out of 100 times it's raining.
00:49:05.490 --> 00:49:08.225
So 1/4 is the probability that it's
00:49:08.225 --> 00:49:11.240
raining times the Entropy of cloudiness
00:49:11.240 --> 00:49:13.840
given that it's raining plus three
00:49:13.840 --> 00:49:15.710
quarter times it's not raining times
00:49:15.710 --> 00:49:17.570
the entropy of the cloudiness given
00:49:17.570 --> 00:49:18.810
that it's not raining.
00:49:20.470 --> 00:49:23.420
So that's a measure of how much does
00:49:23.420 --> 00:49:25.930
knowing whether or not it's rainy, or
00:49:25.930 --> 00:49:28.470
how much uncertainty do I have left if
00:49:28.470 --> 00:49:29.880
I know whether or not it's raining.
00:49:32.430 --> 00:49:34.030
How much do I expect to have left?
00:49:37.700 --> 00:49:39.800
So some useful things to know is that
00:49:39.800 --> 00:49:41.585
the Entropy is always nonnegative.
00:49:41.585 --> 00:49:43.580
You can never have negative Entropy,
00:49:43.580 --> 00:49:45.410
but do make sure you remember.
00:49:46.480 --> 00:49:47.310
00:49:48.750 --> 00:49:50.380
So do make sure you remember these
00:49:50.380 --> 00:49:53.390
negative signs in this like
00:49:53.390 --> 00:49:54.910
probability, otherwise if you end up
00:49:54.910 --> 00:49:56.780
with a negative Entropy that you left
00:49:56.780 --> 00:49:57.490
something out.
00:49:59.760 --> 00:50:02.815
You also have this chain rule, so the
00:50:02.815 --> 00:50:06.320
entropy X&Y is the entropy of X given Y
00:50:06.320 --> 00:50:08.580
plus the entropy of Y, which kind of
00:50:08.580 --> 00:50:10.260
makes sense because the Entropy of
00:50:10.260 --> 00:50:11.280
knowing two things.
00:50:12.310 --> 00:50:14.540
Of the values of two things, is the
00:50:14.540 --> 00:50:15.914
value of knowing one.
00:50:15.914 --> 00:50:18.785
Is the OR sorry, the Entropy or the
00:50:18.785 --> 00:50:20.199
uncertainty of knowing two things?
00:50:20.199 --> 00:50:22.179
Is the uncertainty of knowing one of
00:50:22.180 --> 00:50:22.580
them?
00:50:23.280 --> 00:50:24.940
Plus the uncertainty of knowing the
00:50:24.940 --> 00:50:26.515
other one, given that you already know
00:50:26.515 --> 00:50:27.350
One South.
00:50:27.350 --> 00:50:30.169
It's either Entropy of X given Y plus
00:50:30.169 --> 00:50:32.323
Entropy of Y, or Entropy of Y given X
00:50:32.323 --> 00:50:33.250
plus Entropy of 784x1.
00:50:34.640 --> 00:50:38.739
X&Y are independent, then Entropy of Y
00:50:38.740 --> 00:50:40.659
given X is equal the entropy of Y.
00:50:42.870 --> 00:50:44.520
Meaning that 784X1 doesn't reduce our
00:50:44.520 --> 00:50:45.240
uncertainty at all.
00:50:46.530 --> 00:50:48.845
And Entropy of anything with itself is
00:50:48.845 --> 00:50:50.330
0, because once you know it, then
00:50:50.330 --> 00:50:51.480
there's no uncertainty anymore.
00:50:52.880 --> 00:50:53.390
And then?
00:50:54.110 --> 00:50:57.970
If you do know something, Entropy of Y
00:50:57.970 --> 00:50:59.780
given X or at least has to be less than
00:50:59.780 --> 00:51:01.430
or equal the entropy of Y.
00:51:01.430 --> 00:51:04.020
So knowing something can never increase
00:51:04.020 --> 00:51:04.690
your uncertainty.
00:51:07.660 --> 00:51:09.520
So then finally we can get to this
00:51:09.520 --> 00:51:11.132
information gain.
00:51:11.132 --> 00:51:14.730
So information gain is the change in
00:51:14.730 --> 00:51:17.530
the Entropy due to learning something
00:51:17.530 --> 00:51:17.810
new.
00:51:20.100 --> 00:51:23.310
So I can say, for example, what is?
00:51:23.310 --> 00:51:26.160
How much does knowing whether or not
00:51:26.160 --> 00:51:27.010
it's rainy?
00:51:27.960 --> 00:51:30.610
Reduce my uncertainty of cloudiness.
00:51:31.620 --> 00:51:34.542
So that would be the Entropy of
00:51:34.542 --> 00:51:37.242
cloudiness minus the entropy of
00:51:37.242 --> 00:51:39.120
cloudiness given whether or not it's
00:51:39.120 --> 00:51:39.450
raining.
00:51:41.710 --> 00:51:43.990
So that's the Entropy of cloudiness
00:51:43.990 --> 00:51:46.500
minus the entropy of cloudiness given
00:51:46.500 --> 00:51:47.640
whether it's raining.
00:51:47.640 --> 00:51:49.660
And that's 25 bits.
00:51:49.660 --> 00:51:50.993
So that's like the value.
00:51:50.993 --> 00:51:52.860
It's essentially the value of knowing
00:51:52.860 --> 00:51:54.100
whether or not it's meaning.
00:51:59.210 --> 00:52:01.140
And then finally we can use this in our
00:52:01.140 --> 00:52:02.140
Decision tree.
00:52:02.140 --> 00:52:03.660
So if we recall.
00:52:04.300 --> 00:52:07.310
The Decision tree algorithm is that.
00:52:08.410 --> 00:52:10.940
If I'm trying to I go through like
00:52:10.940 --> 00:52:12.700
splitting my data.
00:52:13.550 --> 00:52:15.050
Choose some Test.
00:52:15.050 --> 00:52:17.280
According to the test, I split the data
00:52:17.280 --> 00:52:18.970
into different nodes and then I choose
00:52:18.970 --> 00:52:20.440
a new test for each of those nodes.
00:52:21.440 --> 00:52:22.840
So the key thing we're trying to figure
00:52:22.840 --> 00:52:24.007
out is how do we do that Test?
00:52:24.007 --> 00:52:25.800
How do we choose the features or
00:52:25.800 --> 00:52:27.480
attributes and the splitting value?
00:52:28.370 --> 00:52:30.100
To try to split things into different
00:52:30.100 --> 00:52:32.030
classes, or in other words, to try to
00:52:32.030 --> 00:52:33.640
reduce the uncertainty of our
00:52:33.640 --> 00:52:34.150
prediction.
00:52:36.190 --> 00:52:39.790
And the solution is to choose the
00:52:39.790 --> 00:52:42.450
attribute to choose the Test that
00:52:42.450 --> 00:52:44.780
maximizes the information gain.
00:52:44.780 --> 00:52:46.770
In other words, that reduces the
00:52:46.770 --> 00:52:49.600
entropy of the most for the current
00:52:49.600 --> 00:52:50.370
data in that node.
00:52:52.000 --> 00:52:52.530
So.
00:52:53.260 --> 00:52:56.478
What you would do is for each for each
00:52:56.478 --> 00:52:58.700
discrete attribute or discrete feature.
00:52:59.630 --> 00:53:02.063
You can compute the information gain of
00:53:02.063 --> 00:53:04.140
using that using that feature.
00:53:04.140 --> 00:53:06.620
So in the case of.
00:53:07.360 --> 00:53:08.280
Go back a bit.
00:53:09.010 --> 00:53:11.670
To this simple true false all right, so
00:53:11.670 --> 00:53:12.650
for example.
00:53:13.650 --> 00:53:15.520
Here I started out with a pretty high
00:53:15.520 --> 00:53:17.550
Entropy, close to one because 5/8 of
00:53:17.550 --> 00:53:18.050
the time.
00:53:18.690 --> 00:53:20.850
The value of Y is true and three it's
00:53:20.850 --> 00:53:21.180
false.
00:53:22.030 --> 00:53:26.620
And so I can say for X1, what's my
00:53:26.620 --> 00:53:28.970
Entropy after X1?
00:53:28.970 --> 00:53:31.020
It's a 5050 chance that it goes either
00:53:31.020 --> 00:53:31.313
way.
00:53:31.313 --> 00:53:34.020
So this will be 50 * 0 because the
00:53:34.020 --> 00:53:36.541
Entropy here is 0 and this will be 50
00:53:36.541 --> 00:53:36.815
times.
00:53:36.815 --> 00:53:38.630
I don't know, one or something,
00:53:38.630 --> 00:53:40.659
whatever that Entropy is, and so this
00:53:40.659 --> 00:53:42.100
Entropy will be really low.
00:53:43.000 --> 00:53:45.700
And this Entropy is just about as high
00:53:45.700 --> 00:53:46.590
as I started with.
00:53:46.590 --> 00:53:48.330
It's only a little bit lower maybe
00:53:48.330 --> 00:53:50.510
because if I go this way, I have
00:53:50.510 --> 00:53:52.691
Entropy of 1, there's a 50% chance of
00:53:52.691 --> 00:53:55.188
that, and if I go this way, then I have
00:53:55.188 --> 00:53:56.721
lower Entropy and there's a 50% chance
00:53:56.721 --> 00:53:57.159
of that.
00:53:57.870 --> 00:54:00.010
And so my information gain is my
00:54:00.010 --> 00:54:01.600
initial entropy of Y.
00:54:02.980 --> 00:54:06.550
Minus the entropy of each of these, and
00:54:06.550 --> 00:54:08.005
here the Entropy gain.
00:54:08.005 --> 00:54:10.210
The information gain of X1 is much
00:54:10.210 --> 00:54:12.940
lower than X2 and so I Choose X1.
00:54:18.810 --> 00:54:20.420
So if I have discrete values, I just
00:54:20.420 --> 00:54:22.449
compute the information gain for the
00:54:22.450 --> 00:54:24.290
current node for each of those discrete
00:54:24.290 --> 00:54:25.725
values, and then I choose the one with
00:54:25.725 --> 00:54:26.860
the highest information gain.
00:54:27.780 --> 00:54:29.650
If I have continuous values, it's
00:54:29.650 --> 00:54:31.576
slightly more complicated because then
00:54:31.576 --> 00:54:34.395
I have to also choose a threshold in
00:54:34.395 --> 00:54:36.230
the lemons and.
00:54:36.920 --> 00:54:40.150
And oranges we were choosing saying if
00:54:40.150 --> 00:54:42.010
the height is greater than six then we
00:54:42.010 --> 00:54:42.620
go one way.
00:54:44.560 --> 00:54:46.640
So we have to choose which feature and
00:54:46.640 --> 00:54:47.420
which threshold.
00:54:48.430 --> 00:54:49.580
So typically.
00:54:51.060 --> 00:54:53.512
Something this I don't know.
00:54:53.512 --> 00:54:56.295
Like who thought putting a projector in
00:54:56.295 --> 00:54:57.930
a jewel would be like a nice way to?
00:54:58.590 --> 00:55:00.260
Right and stuff, but anyway.
00:55:04.700 --> 00:55:06.340
But at least it's something, all right?
00:55:06.340 --> 00:55:08.400
So let's say that I have some feature.
00:55:09.420 --> 00:55:11.910
And I've got like some different
00:55:11.910 --> 00:55:13.240
classes and that feature.
00:55:16.190 --> 00:55:18.560
So what I would do is I would usually
00:55:18.560 --> 00:55:19.950
you would sort the values.
00:55:20.890 --> 00:55:22.440
And you're never going to want to split
00:55:22.440 --> 00:55:24.010
between two of the same class, so I
00:55:24.010 --> 00:55:26.469
would never split between the two X's,
00:55:26.470 --> 00:55:29.250
because that's always going to be worse
00:55:29.250 --> 00:55:31.070
than some split that's between
00:55:31.070 --> 00:55:31.930
different classes.
00:55:32.630 --> 00:55:35.450
So I can consider the thresholds that
00:55:35.450 --> 00:55:35.810
are.
00:55:36.460 --> 00:55:37.730
Between different classes.
00:55:42.380 --> 00:55:44.000
Really.
00:55:44.000 --> 00:55:44.530
No.
00:55:46.130 --> 00:55:48.380
Yeah, I can, but I'm not going to draw
00:55:48.380 --> 00:55:50.220
that long, so it's not worth it to me
00:55:50.220 --> 00:55:50.860
to move on here.
00:55:50.860 --> 00:55:52.160
Then I have to move my laptop and.
00:55:53.030 --> 00:55:56.030
So I'm fine.
00:55:56.750 --> 00:55:59.310
So I would choose these two thresholds.
00:55:59.310 --> 00:56:01.680
If it's this threshold, then it's
00:56:01.680 --> 00:56:04.152
basically two and zero.
00:56:04.152 --> 00:56:07.470
So it's a very low Entropy here.
00:56:07.470 --> 00:56:10.420
And the probability of that is 2 out of
00:56:10.420 --> 00:56:11.505
five, right?
00:56:11.505 --> 00:56:16.820
So it would be 0.4 * 0 is the.
00:56:17.460 --> 00:56:18.580
Entropy on this side.
00:56:19.570 --> 00:56:20.820
And if I go this way?
00:56:21.670 --> 00:56:23.500
Then it's going to be.
00:56:24.440 --> 00:56:25.290
Then I've got.
00:56:26.660 --> 00:56:27.840
Sorry, two out of seven.
00:56:29.750 --> 00:56:31.320
Out of seven times.
00:56:32.470 --> 00:56:33.930
Times Entropy of 0 this way.
00:56:34.650 --> 00:56:37.630
And if I go this way, then it's five
00:56:37.630 --> 00:56:38.020
out of.
00:56:38.980 --> 00:56:39.760
7.
00:56:41.040 --> 00:56:41.770
Times.
00:56:44.510 --> 00:56:47.560
Two out of five times log.
00:56:52.980 --> 00:56:53.690
Thank you.
00:56:53.690 --> 00:56:55.330
I always forget the minus sign.
00:56:56.140 --> 00:56:58.270
OK, so minus 5 to 7, which is a
00:56:58.270 --> 00:56:59.880
probability that I go in this direction
00:56:59.880 --> 00:57:03.805
times one out of five times log one out
00:57:03.805 --> 00:57:04.700
of five.
00:57:05.550 --> 00:57:07.760
Plus four out of five.
00:57:09.170 --> 00:57:10.710
Four to five times log.
00:57:13.360 --> 00:57:14.100
Right.
00:57:14.100 --> 00:57:15.750
So there's a one fifth chance that it's
00:57:15.750 --> 00:57:16.270
an X.
00:57:17.350 --> 00:57:19.180
I do 1/5 times log 1/5.
00:57:19.820 --> 00:57:22.200
Minus 4/5 chance that it's a no, so
00:57:22.200 --> 00:57:23.790
minus 4/5 times log four fifth.
00:57:24.510 --> 00:57:26.210
And this whole thing is the Entropy
00:57:26.210 --> 00:57:27.140
after that split.
00:57:28.590 --> 00:57:30.650
And then likewise I can evaluate this
00:57:30.650 --> 00:57:32.850
split as well and so.
00:57:33.620 --> 00:57:35.650
Out of these two splits, which one do
00:57:35.650 --> 00:57:37.190
you think will have the most
00:57:37.190 --> 00:57:38.040
information gain?
00:57:41.220 --> 00:57:43.320
Yeah, the left split, the first one has
00:57:43.320 --> 00:57:45.050
the most information gain because then
00:57:45.050 --> 00:57:47.168
I get a confident Decision about two
00:57:47.168 --> 00:57:49.943
X's and like 4 out of five chance of
00:57:49.943 --> 00:57:51.739
getting it right on the other side,
00:57:51.740 --> 00:57:53.520
where if I choose the right split, I
00:57:53.520 --> 00:57:56.791
only get a perfect confidence about 1X
00:57:56.791 --> 00:57:59.529
and A2 out of three chance of getting
00:57:59.529 --> 00:58:00.529
it right on the other side.
00:58:15.920 --> 00:58:19.580
OK, so if I continuous features I would
00:58:19.580 --> 00:58:21.490
just try all the different like
00:58:21.490 --> 00:58:23.110
candidate thresholds for all those
00:58:23.110 --> 00:58:24.690
features and then choose the best one.
00:58:26.430 --> 00:58:28.360
And.
00:58:28.460 --> 00:58:29.720
She's the best one, all right.
00:58:29.720 --> 00:58:30.090
That's it.
00:58:30.090 --> 00:58:31.430
And then I do that for all the nodes,
00:58:31.430 --> 00:58:32.590
then I do it Recursively.
00:58:33.670 --> 00:58:35.660
So if you have a lot of features and a
00:58:35.660 --> 00:58:37.050
lot of data, this can kind of take a
00:58:37.050 --> 00:58:37.600
long time.
00:58:38.250 --> 00:58:40.610
But I mean these operations are super
00:58:40.610 --> 00:58:41.710
fast so.
00:58:42.980 --> 00:58:45.919
In practice, when you run it so in
00:58:45.920 --> 00:58:48.380
homework two, I'll have you train tree
00:58:48.380 --> 00:58:50.930
train forests of Decision trees, where
00:58:50.930 --> 00:58:54.165
you train 100 of them for example, and
00:58:54.165 --> 00:58:56.090
it takes like a few seconds, so it's
00:58:56.090 --> 00:58:57.204
like pretty fast.
00:58:57.204 --> 00:58:59.070
These are these are actually not that
00:58:59.070 --> 00:59:01.030
computationally expensive, even though
00:59:01.030 --> 00:59:02.610
doing it manually would take forever.
00:59:05.590 --> 00:59:06.980
So.
00:59:08.860 --> 00:59:10.970
We're close to the we're close to the
00:59:10.970 --> 00:59:11.690
end of the lecture.
00:59:12.320 --> 00:59:14.320
But I will give you just a second to
00:59:14.320 --> 00:59:15.230
catch your breath.
00:59:15.230 --> 00:59:17.030
And while you're doing that, think
00:59:17.030 --> 00:59:17.690
about.
00:59:19.060 --> 00:59:22.640
If I were to try and in this case I'm
00:59:22.640 --> 00:59:23.760
showing like all the different
00:59:23.760 --> 00:59:25.210
examples, the numbers are different
00:59:25.210 --> 00:59:27.530
examples there and the color is whether
00:59:27.530 --> 00:59:28.270
they wait or not.
00:59:28.850 --> 00:59:30.570
And I'm trying to decide whether I'm
00:59:30.570 --> 00:59:33.090
going to make a decision based on the
00:59:33.090 --> 00:59:35.096
type of restaurant or based on whether
00:59:35.096 --> 00:59:35.860
the restaurant's full.
00:59:36.490 --> 00:59:40.840
So take a moment to stretch or zone
00:59:40.840 --> 00:59:42.760
out, and then I'll ask you what the
00:59:42.760 --> 00:59:43.200
answer is.
01:00:05.270 --> 01:00:06.606
Part of it, yeah.
01:00:06.606 --> 01:00:08.755
So this is all Training one tree.
01:00:08.755 --> 01:00:10.840
And for a random forest you just
01:00:10.840 --> 01:00:14.246
randomly sample features and randomly
01:00:14.246 --> 01:00:16.760
sample data, and then you train a tree
01:00:16.760 --> 01:00:19.250
and then you do that like N times and
01:00:19.250 --> 01:00:20.600
then you average the predictions.
01:00:27.420 --> 01:00:27.810
Yeah.
01:00:30.860 --> 01:00:33.610
And so essentially, since the previous
01:00:33.610 --> 01:00:35.440
Entropy is fixed when you're trying to
01:00:35.440 --> 01:00:36.140
make a decision.
01:00:36.910 --> 01:00:38.739
You're just essentially choosing the
01:00:38.740 --> 01:00:41.810
Decision, choosing the attribute that
01:00:41.810 --> 01:00:45.320
will minimize your expected Entropy
01:00:45.320 --> 01:00:47.160
after, like given that attribute.
01:00:57.950 --> 01:01:00.790
Alright, so how many people think that
01:01:00.790 --> 01:01:02.610
we should split?
01:01:03.300 --> 01:01:04.710
How many people think we should split
01:01:04.710 --> 01:01:05.580
based on type?
01:01:08.180 --> 01:01:09.580
How many people think we should split
01:01:09.580 --> 01:01:10.520
based on Patrons?
01:01:12.730 --> 01:01:13.680
Yeah, OK.
01:01:14.430 --> 01:01:17.380
So I would say the answer is Patrons
01:01:17.380 --> 01:01:19.870
and because splitting based on type.
01:01:20.590 --> 01:01:22.310
I end up no matter what type of
01:01:22.310 --> 01:01:24.200
restaurant is, I end up with an equal
01:01:24.200 --> 01:01:26.120
number of greens and Reds.
01:01:26.120 --> 01:01:30.140
So green green means I didn't like say
01:01:30.140 --> 01:01:32.672
it very clearly, but green means that
01:01:32.672 --> 01:01:35.842
you think that you go, that you wait,
01:01:35.842 --> 01:01:37.540
and red means that you don't wait.
01:01:38.460 --> 01:01:40.820
So type tells me nothing, right?
01:01:40.820 --> 01:01:42.310
It doesn't help me split anything at
01:01:42.310 --> 01:01:42.455
all.
01:01:42.455 --> 01:01:44.898
I knew initially I had complete Entropy
01:01:44.898 --> 01:01:47.780
Entropy of 1 and after knowing type I
01:01:47.780 --> 01:01:48.880
still have Entropy of 1.
01:01:49.900 --> 01:01:52.140
Where if I know Patrons, then a lot of
01:01:52.140 --> 01:01:55.720
the time I have my Decision, and only
01:01:55.720 --> 01:01:57.230
some fraction of the time I still have
01:01:57.230 --> 01:01:57.590
to.
01:01:57.590 --> 01:01:59.040
I need more information.
01:02:00.990 --> 01:02:02.700
So here's like all the math.
01:02:04.250 --> 01:02:05.230
To go through that but.
01:02:08.910 --> 01:02:11.790
All right, So what if I?
01:02:12.780 --> 01:02:14.730
So sometimes a lot of times trees are
01:02:14.730 --> 01:02:16.930
used for continuous values and then
01:02:16.930 --> 01:02:18.320
it's called a Regression tree.
01:02:20.760 --> 01:02:22.960
The Regression tree is learned in the
01:02:22.960 --> 01:02:23.510
same way.
01:02:24.570 --> 01:02:29.490
Except that you would use the instead
01:02:29.490 --> 01:02:30.840
of, sorry.
01:02:32.260 --> 01:02:34.170
In the Regression tree, it's the same
01:02:34.170 --> 01:02:36.530
way, but you're typically trying to
01:02:36.530 --> 01:02:38.703
minimize the sum of squared error of
01:02:38.703 --> 01:02:41.862
the node instead of minimizing the
01:02:41.862 --> 01:02:42.427
cross entropy.
01:02:42.427 --> 01:02:44.050
You could still do it actually based on
01:02:44.050 --> 01:02:45.500
cross entropy if you're seeing like
01:02:45.500 --> 01:02:47.770
Gaussian distributions, but here let me
01:02:47.770 --> 01:02:48.900
show you an example.
01:02:54.540 --> 01:02:55.230
So.
01:02:57.600 --> 01:02:59.170
Let's just say I'm doing like one
01:02:59.170 --> 01:03:00.700
feature, let's say like.
01:03:01.480 --> 01:03:04.610
This is my feature X and my prediction
01:03:04.610 --> 01:03:05.240
value.
01:03:06.000 --> 01:03:07.830
Is the number that I'm putting here.
01:03:18.340 --> 01:03:18.690
OK.
01:03:19.430 --> 01:03:20.160
So.
01:03:21.350 --> 01:03:22.980
I'm trying to predict what this number
01:03:22.980 --> 01:03:25.460
is given like where I fell on this X
01:03:25.460 --> 01:03:25.950
axis.
01:03:27.030 --> 01:03:28.940
So the best split I could do is
01:03:28.940 --> 01:03:30.550
probably like here, right?
01:03:31.330 --> 01:03:33.800
And if I take this split, then I would
01:03:33.800 --> 01:03:37.313
say that if I'm in this side of the
01:03:37.313 --> 01:03:37.879
split.
01:03:38.640 --> 01:03:42.450
Then my prediction is 4 out of three,
01:03:42.450 --> 01:03:44.560
which is the average of the values that
01:03:44.560 --> 01:03:45.690
are on this side of the split.
01:03:46.510 --> 01:03:48.710
And if I'm on this side of the split,
01:03:48.710 --> 01:03:51.030
then my prediction is 6.
01:03:51.800 --> 01:03:53.900
Which is 18 / 3, right?
01:03:53.900 --> 01:03:55.815
So it's the average of these values.
01:03:55.815 --> 01:03:58.270
So if I'm doing Regression, I'm still
01:03:58.270 --> 01:04:00.520
like I'm choosing a split that's going
01:04:00.520 --> 01:04:03.580
to give me the best prediction in each
01:04:03.580 --> 01:04:04.390
side of the split.
01:04:04.980 --> 01:04:06.745
And then my estimate on each side of
01:04:06.745 --> 01:04:08.170
the split is just the average of the
01:04:08.170 --> 01:04:10.100
values after that split.
01:04:11.000 --> 01:04:13.950
And the scoring, the scoring that I can
01:04:13.950 --> 01:04:16.120
use is the squared error.
01:04:16.120 --> 01:04:20.464
So if the squared error would be 1 -, 4
01:04:20.464 --> 01:04:22.536
thirds squared, plus 2 -, 4 thirds
01:04:22.536 --> 01:04:24.905
squared plus 1 -, 4 thirds squared plus
01:04:24.905 --> 01:04:29.049
5 -, 6 ^2 + 8 -, 6 ^2 + 5 -, 6 ^2.
01:04:29.890 --> 01:04:31.665
And so I could try like every
01:04:31.665 --> 01:04:33.549
threshold, compute my squared error
01:04:33.550 --> 01:04:35.635
given every threshold and then choose
01:04:35.635 --> 01:04:37.060
the one that gives me the lowest
01:04:37.060 --> 01:04:37.740
squared error.
01:04:41.040 --> 01:04:43.055
So it's the same algorithm, except that
01:04:43.055 --> 01:04:44.245
you have a different Criterion.
01:04:44.245 --> 01:04:46.370
You might use squared error.
01:04:47.820 --> 01:04:49.530
Because it's continuous values that I'm
01:04:49.530 --> 01:04:50.100
predicting.
01:04:50.840 --> 01:04:53.030
And then the output of the node will be
01:04:53.030 --> 01:04:54.390
the average of the Samples that fall
01:04:54.390 --> 01:04:55.069
into that node.
01:04:56.480 --> 01:04:58.300
And for Regression trees, that's
01:04:58.300 --> 01:05:00.020
especially important to.
01:05:01.330 --> 01:05:03.490
Stop growing your tree early, because
01:05:03.490 --> 01:05:05.420
obviously otherwise you're going to
01:05:05.420 --> 01:05:10.090
always separate your data into one leaf
01:05:10.090 --> 01:05:12.030
node per data point, since you have
01:05:12.030 --> 01:05:13.995
like continuous values, unless there's
01:05:13.995 --> 01:05:15.410
like many of the same value.
01:05:16.020 --> 01:05:17.410
And so you're going to tend to like
01:05:17.410 --> 01:05:17.870
overfit.
01:05:23.330 --> 01:05:25.920
Overfitting, by the way, that's a term
01:05:25.920 --> 01:05:27.060
that comes up a lot in machine
01:05:27.060 --> 01:05:27.865
learning.
01:05:27.865 --> 01:05:30.870
Overfitting means that your model you
01:05:30.870 --> 01:05:32.910
have a very complex model so that you
01:05:32.910 --> 01:05:34.940
achieve like really low Training Error.
01:05:35.660 --> 01:05:37.550
But due to the complexity you're Test
01:05:37.550 --> 01:05:38.740
error has gone up.
01:05:38.740 --> 01:05:41.570
So if you plot your.
01:05:42.250 --> 01:05:44.210
If you plot your Test error as, you
01:05:44.210 --> 01:05:45.510
increase complexity.
01:05:46.200 --> 01:05:48.000
You're Test error will go down for some
01:05:48.000 --> 01:05:50.030
time, but then at some point as your
01:05:50.030 --> 01:05:51.880
complexity keeps rising, you're Test
01:05:51.880 --> 01:05:53.590
Error will start to increase.
01:05:53.590 --> 01:05:55.040
So the point at which you're.
01:05:55.740 --> 01:05:57.650
You're Test Error increases due to
01:05:57.650 --> 01:05:59.260
increasing complexity is where you
01:05:59.260 --> 01:06:00.040
start overfitting.
01:06:00.870 --> 01:06:02.300
We'll talk about that more at the start
01:06:02.300 --> 01:06:03.000
of the ensembles.
01:06:04.840 --> 01:06:06.610
Right, so there's a few variants.
01:06:06.610 --> 01:06:08.620
You can use different splitting
01:06:08.620 --> 01:06:09.490
criteria.
01:06:09.490 --> 01:06:12.010
For example, the genie like impurity or
01:06:12.010 --> 01:06:14.580
Genie Diversity index is just one minus
01:06:14.580 --> 01:06:17.460
the sum over all the values of X
01:06:17.460 --> 01:06:18.700
probability of X ^2.
01:06:19.480 --> 01:06:22.140
This actually is like almost the same
01:06:22.140 --> 01:06:25.800
thing as the Entropy.
01:06:26.410 --> 01:06:27.800
But it's a little bit faster to
01:06:27.800 --> 01:06:29.840
compute, so it's actually more often
01:06:29.840 --> 01:06:30.740
used as the default.
01:06:33.830 --> 01:06:35.890
Most times you split on one attribute
01:06:35.890 --> 01:06:38.460
at a time, but you can also.
01:06:39.190 --> 01:06:40.820
They're in some algorithms.
01:06:40.820 --> 01:06:42.790
You can solve for slices through the
01:06:42.790 --> 01:06:44.600
feature space you can.
01:06:45.280 --> 01:06:47.490
Do like linear discriminant analysis or
01:06:47.490 --> 01:06:49.200
something like that to try to find like
01:06:49.200 --> 01:06:51.970
a multivariable split that separates
01:06:51.970 --> 01:06:53.870
the data, but usually it's just single
01:06:53.870 --> 01:06:54.310
attribute.
01:06:56.180 --> 01:06:57.970
And as I mentioned a couple of times,
01:06:57.970 --> 01:07:00.010
you can stop early so you don't need to
01:07:00.010 --> 01:07:02.010
grow like the full tree until you get
01:07:02.010 --> 01:07:03.025
perfect Training accuracy.
01:07:03.025 --> 01:07:06.110
You can stop after you reach a Max
01:07:06.110 --> 01:07:09.475
depth or stop after you have a certain
01:07:09.475 --> 01:07:11.540
number of nodes per certain number of
01:07:11.540 --> 01:07:12.620
data points per node.
01:07:13.710 --> 01:07:15.470
And the reason that you had stopped
01:07:15.470 --> 01:07:16.920
early is because you the tree to
01:07:16.920 --> 01:07:18.990
generalized new data and if you grow
01:07:18.990 --> 01:07:20.450
like a really big tree, you're going to
01:07:20.450 --> 01:07:23.000
end up with these like little like
01:07:23.000 --> 01:07:26.240
micro applicable rules that might not
01:07:26.240 --> 01:07:27.750
work well when you get new Test
01:07:27.750 --> 01:07:28.270
Samples.
01:07:29.220 --> 01:07:31.190
Where if you have a shorter tree that
01:07:31.190 --> 01:07:34.260
then you might have some uncertainty
01:07:34.260 --> 01:07:36.147
left in your leaf nodes, but you can
01:07:36.147 --> 01:07:38.300
have more confidence that will reflect
01:07:38.300 --> 01:07:39.240
the true distribution.
01:07:42.350 --> 01:07:45.630
So if we look at Decision trees versus
01:07:45.630 --> 01:07:46.280
one and north.
01:07:46.980 --> 01:07:49.500
They're actually kind of similar in a
01:07:49.500 --> 01:07:49.950
way.
01:07:49.950 --> 01:07:51.620
They both have piecewise linear
01:07:51.620 --> 01:07:52.120
decisions.
01:07:52.750 --> 01:07:54.620
So here's the boundary that I get with
01:07:54.620 --> 01:07:56.420
one and N in this example.
01:07:57.110 --> 01:08:00.550
It's going to be based on like if you
01:08:00.550 --> 01:08:03.380
chop things up into cells where each
01:08:03.380 --> 01:08:05.770
sample is like everything within the
01:08:05.770 --> 01:08:07.550
cell is closest to a particular sample.
01:08:08.260 --> 01:08:09.460
I would get this boundary.
01:08:11.100 --> 01:08:12.915
And with the Decision tree you tend to
01:08:12.915 --> 01:08:14.440
get, if you're doing 1 attribute at a
01:08:14.440 --> 01:08:15.980
time, you get this access to line
01:08:15.980 --> 01:08:16.630
boundary.
01:08:16.630 --> 01:08:18.832
So it ends up being like going straight
01:08:18.832 --> 01:08:20.453
over and then up and then straight over
01:08:20.453 --> 01:08:22.160
and then down and then a little bit
01:08:22.160 --> 01:08:23.320
over and then down.
01:08:23.320 --> 01:08:25.226
But they're kind of similar.
01:08:25.226 --> 01:08:28.220
So they're the overlap of those spaces
01:08:28.220 --> 01:08:28.690
is similar.
01:08:31.900 --> 01:08:34.170
The Decision tree also has the ability
01:08:34.170 --> 01:08:36.042
for over stopping to improve
01:08:36.042 --> 01:08:36.520
generalization.
01:08:36.520 --> 01:08:38.530
While they can and doesn't the K&N you
01:08:38.530 --> 01:08:40.700
can increase K to try to improve
01:08:40.700 --> 01:08:42.110
generalization to make it like a
01:08:42.110 --> 01:08:44.506
smoother boundary, but it doesn't have
01:08:44.506 --> 01:08:46.540
like as doesn't have very many like
01:08:46.540 --> 01:08:47.930
controls or knobs to tune.
01:08:50.390 --> 01:08:53.010
And the true power that Decision trees
01:08:53.010 --> 01:08:54.580
arise with ensembles.
01:08:54.580 --> 01:08:56.920
So if you combine lots of these trees
01:08:56.920 --> 01:08:59.250
together to make a prediction, then
01:08:59.250 --> 01:09:01.050
suddenly it becomes very effective.
01:09:01.750 --> 01:09:04.430
In practice, people don't usually use
01:09:04.430 --> 01:09:06.620
this one Decision tree in machine
01:09:06.620 --> 01:09:07.998
learning to make an automated
01:09:07.998 --> 01:09:08.396
prediction.
01:09:08.396 --> 01:09:10.710
They usually use a whole bunch of them
01:09:10.710 --> 01:09:12.397
and then average the results or train
01:09:12.397 --> 01:09:14.870
them in a way that they that they
01:09:14.870 --> 01:09:17.126
incrementally build up your prediction.
01:09:17.126 --> 01:09:18.850
And that's what I'll talk about when I
01:09:18.850 --> 01:09:19.730
talk about ensembles.
01:09:22.360 --> 01:09:23.750
So Decision trees are really a
01:09:23.750 --> 01:09:26.740
component in two of the most successful
01:09:26.740 --> 01:09:28.970
algorithms of all time, but they're not
01:09:28.970 --> 01:09:29.630
the whole thing.
01:09:30.940 --> 01:09:33.160
Here's an example of a Regression tree
01:09:33.160 --> 01:09:34.470
for Temperature prediction.
01:09:35.560 --> 01:09:37.200
Just so that I can make the tree simple
01:09:37.200 --> 01:09:39.370
enough to put on a Slide, I set the Min
01:09:39.370 --> 01:09:41.840
leaf size to 200 so there.
01:09:41.840 --> 01:09:44.000
So I stopped splitting once the node
01:09:44.000 --> 01:09:44.990
has 200 points.
01:09:46.120 --> 01:09:49.080
And then I computed the root mean
01:09:49.080 --> 01:09:50.450
squared error and the R2.
01:09:51.680 --> 01:09:53.280
And so you can see for example like.
01:09:54.430 --> 01:09:55.990
One thing that is interesting to me
01:09:55.990 --> 01:09:58.278
about this is that I would have thought
01:09:58.278 --> 01:09:59.872
that the temperature in Cleveland
01:09:59.872 --> 01:10:01.510
yesterday would be the best predictor
01:10:01.510 --> 01:10:03.150
of the temperature in Cleveland today,
01:10:03.150 --> 01:10:05.056
but it's actually not the best
01:10:05.056 --> 01:10:05.469
predictor.
01:10:05.470 --> 01:10:09.090
So the best single like criteria is the
01:10:09.090 --> 01:10:11.090
temperature in Chicago yesterday,
01:10:11.090 --> 01:10:13.590
because I guess the weather like moves
01:10:13.590 --> 01:10:15.460
from West to east a bit.
01:10:16.850 --> 01:10:19.665
And I guess downward, so knowing the
01:10:19.665 --> 01:10:21.477
weather in Chicago yesterday, whether
01:10:21.477 --> 01:10:23.420
the weather was less than whether the
01:10:23.420 --> 01:10:25.530
Temperature was less than 8.4 Celsius
01:10:25.530 --> 01:10:26.950
or greater than 8.4 Celsius.
01:10:27.590 --> 01:10:29.040
Is the best single thing that I can
01:10:29.040 --> 01:10:29.290
know.
01:10:30.480 --> 01:10:31.160
And then?
01:10:32.290 --> 01:10:34.920
That reduces my initial squared error
01:10:34.920 --> 01:10:36.400
was 112.
01:10:38.170 --> 01:10:39.680
And then if you divide it by number of
01:10:39.680 --> 01:10:41.560
Samples, then or.
01:10:42.810 --> 01:10:44.720
Yeah, take divided by number of samples
01:10:44.720 --> 01:10:45.960
and take square root or something to
01:10:45.960 --> 01:10:47.390
get the per sample.
01:10:48.300 --> 01:10:51.010
Then depending on that answer, then I
01:10:51.010 --> 01:10:53.209
check to see what is the temperature in
01:10:53.210 --> 01:10:55.458
Milwaukee yesterday or what is the
01:10:55.458 --> 01:10:57.140
temperature in Grand Rapids yesterday.
01:10:58.060 --> 01:10:59.600
And then depending on those answers, I
01:10:59.600 --> 01:11:02.040
check Chicago again, a different value
01:11:02.040 --> 01:11:04.170
of Chicago, and then I get my final
01:11:04.170 --> 01:11:04.840
decision here.
01:11:11.120 --> 01:11:13.720
Yeah, it's like my sister lives in
01:11:13.720 --> 01:11:16.140
Harrisburg, so I always know that
01:11:16.140 --> 01:11:17.750
they're going to get our weather like a
01:11:17.750 --> 01:11:18.240
day later.
01:11:19.020 --> 01:11:20.680
So it's like, it's really warm here.
01:11:20.680 --> 01:11:22.190
They're like, it's cold, it's warm
01:11:22.190 --> 01:11:22.450
here.
01:11:22.450 --> 01:11:23.600
Well, I guess it will be warm for you
01:11:23.600 --> 01:11:24.930
tomorrow or in two days.
01:11:26.130 --> 01:11:26.620
Yeah.
01:11:27.540 --> 01:11:29.772
But part of the reason that I share
01:11:29.772 --> 01:11:31.300
this is that the one thing that's
01:11:31.300 --> 01:11:32.890
really cool about Decision trees is
01:11:32.890 --> 01:11:34.910
that you get some explanation, like you
01:11:34.910 --> 01:11:37.450
can understand the data better by
01:11:37.450 --> 01:11:39.600
looking at the tree like this kind of
01:11:39.600 --> 01:11:41.860
violated my initial assumption that the
01:11:41.860 --> 01:11:43.340
best thing to know for the Temperature
01:11:43.340 --> 01:11:44.830
is your Temperature the previous day.
01:11:45.460 --> 01:11:46.600
It's actually the temperature of
01:11:46.600 --> 01:11:48.965
another city the previous day and you
01:11:48.965 --> 01:11:51.100
can get you can create these rules that
01:11:51.100 --> 01:11:53.030
help you understand, like how to make
01:11:53.030 --> 01:11:53.740
predictions.
01:11:56.130 --> 01:11:56.710
This is.
01:11:56.710 --> 01:11:58.320
I'm not expecting you to read this now,
01:11:58.320 --> 01:12:00.370
but this is the code to generate this
01:12:00.370 --> 01:12:00.820
tree.
01:12:06.080 --> 01:12:07.600
Right on Summary.
01:12:08.580 --> 01:12:10.800
The key assumptions of this of the
01:12:10.800 --> 01:12:12.570
Classification or Regression trees are
01:12:12.570 --> 01:12:15.255
that Samples with similar features have
01:12:15.255 --> 01:12:16.070
similar predictions.
01:12:16.070 --> 01:12:17.590
So it's a similar assumption in Nearest
01:12:17.590 --> 01:12:19.580
neighbor, except this time we're trying
01:12:19.580 --> 01:12:21.680
to figure out how to like split up the
01:12:21.680 --> 01:12:23.420
feature space to define that
01:12:23.420 --> 01:12:25.159
similarity, rather than using like a
01:12:25.160 --> 01:12:29.090
preset distance function like Euclidean
01:12:29.090 --> 01:12:29.560
distance.
01:12:30.970 --> 01:12:32.610
The model parameters are the split
01:12:32.610 --> 01:12:34.560
criteria, each internal node, and then
01:12:34.560 --> 01:12:36.630
the final prediction at each leaf node.
01:12:38.200 --> 01:12:40.020
The designs are putting limits on the
01:12:40.020 --> 01:12:42.080
tree growth and what kinds of splits
01:12:42.080 --> 01:12:43.545
you can consider, like whether to split
01:12:43.545 --> 01:12:45.260
on one attribute or whole groups of
01:12:45.260 --> 01:12:48.930
attributes and then choosing their
01:12:48.930 --> 01:12:50.030
criteria for this split.
01:12:51.520 --> 01:12:52.120
01:12:53.300 --> 01:12:56.060
You Decision trees by themselves are
01:12:56.060 --> 01:12:57.645
useful if you want some explainable
01:12:57.645 --> 01:12:58.710
Decision function.
01:12:58.710 --> 01:13:00.270
So they could be used for like medical
01:13:00.270 --> 01:13:02.090
diagnosis for example, because you want
01:13:02.090 --> 01:13:03.750
to be able to tell people like why.
01:13:04.710 --> 01:13:07.324
Like why I know you have cancer, like
01:13:07.324 --> 01:13:08.840
you don't want to just be like I use
01:13:08.840 --> 01:13:10.270
this machine learning algorithm and it
01:13:10.270 --> 01:13:12.070
says you have like a 93% chance of
01:13:12.070 --> 01:13:13.750
having cancer and so sorry.
01:13:15.000 --> 01:13:16.785
You want to be able to say like because
01:13:16.785 --> 01:13:19.086
of like this thing and because of this
01:13:19.086 --> 01:13:21.099
thing and because of this thing like
01:13:21.100 --> 01:13:24.919
out of all these 1500 cases like 90% of
01:13:24.920 --> 01:13:26.750
them ended up having cancer.
01:13:26.750 --> 01:13:28.180
So we need to do, we need to do a
01:13:28.180 --> 01:13:29.110
biopsy, right.
01:13:29.110 --> 01:13:30.305
So you want some explanation.
01:13:30.305 --> 01:13:31.900
A lot of times it's not always good
01:13:31.900 --> 01:13:33.600
enough to have like a good prediction.
01:13:35.590 --> 01:13:37.240
And they're also like really effective
01:13:37.240 --> 01:13:38.520
as part of a ensemble.
01:13:38.520 --> 01:13:39.650
And again, I think we might see a
01:13:39.650 --> 01:13:40.650
Tuesday instead of Thursday.
01:13:43.150 --> 01:13:44.960
It's not like a really good predictor
01:13:44.960 --> 01:13:47.320
by itself, but it is really good as
01:13:47.320 --> 01:13:47.770
part of an.
01:13:48.500 --> 01:13:48.790
Alright.
01:13:49.670 --> 01:13:51.250
So things you remember, Decision
01:13:51.250 --> 01:13:52.690
Regression trees learn to split up the
01:13:52.690 --> 01:13:54.590
feature space into partitions into
01:13:54.590 --> 01:13:56.110
different cells with similar values.
01:13:57.150 --> 01:13:59.070
And then Entropy is a really important
01:13:59.070 --> 01:13:59.600
concept.
01:13:59.600 --> 01:14:01.030
It's a measure of uncertainty.
01:14:02.730 --> 01:14:05.170
Information gain measures how much
01:14:05.170 --> 01:14:07.090
particular knowledge reduces the
01:14:07.090 --> 01:14:08.710
prediction uncertainty, and that's the
01:14:08.710 --> 01:14:10.260
basis for forming our tree.
01:14:11.630 --> 01:14:13.650
So on Thursday I'm going to do a bit of
01:14:13.650 --> 01:14:15.680
review of our concepts and then I think
01:14:15.680 --> 01:14:17.200
most likely next Tuesday I'll talk
01:14:17.200 --> 01:14:19.730
about ensembles and random forests and
01:14:19.730 --> 01:14:21.540
give you an extensive example of how
01:14:21.540 --> 01:14:23.560
it's used in the Kinect algorithm.
01:14:24.800 --> 01:14:25.770
Alright, thanks everyone.
01:14:25.770 --> 01:14:26.660
See you Thursday.
01:19:07.020 --> 01:19:08.790
Hello.
01:19:10.510 --> 01:19:11.430
Training an assault.