whisper-finetuning-for-asee / CS_441_2023_Spring_February_02,_2023.vtt
classen3's picture
Imported CS 441 audio/transcripts
a67be9a verified
WEBVTT Kind: captions; Language: en-US
NOTE
Created on 2024-02-07T20:53:55.0930204Z by ClassTranscribe
00:01:12.090 --> 00:01:13.230
Alright, good morning everybody.
00:01:15.530 --> 00:01:20.650
So I saw in response to the feedback, I
00:01:20.650 --> 00:01:22.790
got some feedback on the course and.
00:01:23.690 --> 00:01:26.200
Overall, there's of course a mix of
00:01:26.200 --> 00:01:28.650
responses, but some on average people
00:01:28.650 --> 00:01:30.810
feel like it's moving a little fast and
00:01:30.810 --> 00:01:33.040
we're and also it's challenging.
00:01:33.980 --> 00:01:37.350
So I wanted to take some time to like
00:01:37.350 --> 00:01:39.430
consolidate and to talk about some of
00:01:39.430 --> 00:01:40.890
the most important points.
00:01:41.790 --> 00:01:45.160
That we've covered so far, and then so
00:01:45.160 --> 00:01:46.930
I'll do that for the first half of the
00:01:46.930 --> 00:01:49.280
lecture, and then I'm also going to go
00:01:49.280 --> 00:01:53.105
through a detailed example using code
00:01:53.105 --> 00:01:55.200
to solve a particular problem.
00:01:59.460 --> 00:02:00.000
So.
00:02:00.690 --> 00:02:01.340
Let me see.
00:02:04.270 --> 00:02:04.800
All right.
00:02:06.140 --> 00:02:09.360
So this is a mostly the same as a slide
00:02:09.360 --> 00:02:10.750
that I showed in the intro.
00:02:10.750 --> 00:02:12.750
This is machine learning in general.
00:02:12.750 --> 00:02:15.800
You've got some raw features and so far
00:02:15.800 --> 00:02:17.430
we've COVID cases where we have
00:02:17.430 --> 00:02:20.190
discrete and continuous values and also
00:02:20.190 --> 00:02:21.780
some simple images in terms of the
00:02:21.780 --> 00:02:22.670
amnesty characters.
00:02:23.600 --> 00:02:25.590
And we have some kind of.
00:02:25.590 --> 00:02:28.000
Sometimes we process those features in
00:02:28.000 --> 00:02:29.740
some way we have like what's called an
00:02:29.740 --> 00:02:31.970
encoder or we have feature transforms.
00:02:32.790 --> 00:02:34.290
We've only gotten into that a little
00:02:34.290 --> 00:02:34.500
bit.
00:02:35.210 --> 00:02:38.410
In terms of the decision trees, which
00:02:38.410 --> 00:02:39.670
you can view as a kind of feature
00:02:39.670 --> 00:02:40.530
transformation.
00:02:41.290 --> 00:02:43.440
And feature selection using one the
00:02:43.440 --> 00:02:44.350
district regression.
00:02:44.980 --> 00:02:47.030
So the job of the encoder is to take
00:02:47.030 --> 00:02:48.690
your raw features and turn them into
00:02:48.690 --> 00:02:50.570
something that's more easily.
00:02:51.340 --> 00:02:53.270
That more easily yields a predictor.
00:02:54.580 --> 00:02:56.180
Then you have decoder, the thing that
00:02:56.180 --> 00:02:58.290
predicts from your encoded features,
00:02:58.290 --> 00:03:00.510
and we've covered pretty much all the
00:03:00.510 --> 00:03:02.550
methods here except for SVM, which
00:03:02.550 --> 00:03:04.910
we're doing next week.
00:03:05.830 --> 00:03:08.110
And so we've got a linear aggressor, a
00:03:08.110 --> 00:03:09.952
logistic regressor, nearest neighbor
00:03:09.952 --> 00:03:11.430
and probabilistic models.
00:03:11.430 --> 00:03:13.035
Now there's lots of different kinds of
00:03:13.035 --> 00:03:13.650
probabilistic models.
00:03:13.650 --> 00:03:15.930
We only talked about a couple of one of
00:03:15.930 --> 00:03:17.420
them nibs.
00:03:18.750 --> 00:03:20.350
But still, we've touched on this.
00:03:21.360 --> 00:03:22.690
And then you have a prediction and
00:03:22.690 --> 00:03:24.047
there's lots of different things you
00:03:24.047 --> 00:03:24.585
can predict.
00:03:24.585 --> 00:03:26.480
You can predict a category or a
00:03:26.480 --> 00:03:28.050
continuous value, which is what we've
00:03:28.050 --> 00:03:29.205
talked about South far.
00:03:29.205 --> 00:03:31.420
You could also be generating clusters
00:03:31.420 --> 00:03:35.595
or pixel labels or poses or other kinds
00:03:35.595 --> 00:03:36.460
of predictions.
00:03:37.600 --> 00:03:40.275
And in training, you've got some data
00:03:40.275 --> 00:03:42.280
and target labels, and you're trying to
00:03:42.280 --> 00:03:44.060
update the models of your parameters to
00:03:44.060 --> 00:03:46.200
get the best prediction possible, where
00:03:46.200 --> 00:03:48.434
you want to really not only maximize
00:03:48.434 --> 00:03:50.619
your prediction on the training data,
00:03:50.620 --> 00:03:52.970
but also to maximize your expected or
00:03:52.970 --> 00:03:55.520
minimize your expected error on the
00:03:55.520 --> 00:03:56.170
test data.
00:03:59.950 --> 00:04:02.520
So one important part of machine
00:04:02.520 --> 00:04:04.255
learning is learning a model.
00:04:04.255 --> 00:04:04.650
So.
00:04:05.430 --> 00:04:08.385
Here this is like this kind of.
00:04:08.385 --> 00:04:10.610
This function, in one form or another
00:04:10.610 --> 00:04:12.140
will be part of every machine learning
00:04:12.140 --> 00:04:14.180
algorithm where you're trying to.
00:04:14.180 --> 00:04:17.720
You have some model F of X Theta.
00:04:18.360 --> 00:04:20.500
Where X is the raw features.
00:04:21.890 --> 00:04:23.600
Beta are the parameters that you're
00:04:23.600 --> 00:04:25.440
trying to optimize that you're going to
00:04:25.440 --> 00:04:26.840
optimize to fit your model.
00:04:27.940 --> 00:04:31.080
And why is the prediction that you're
00:04:31.080 --> 00:04:31.750
trying to make?
00:04:31.750 --> 00:04:33.359
So you're given.
00:04:33.360 --> 00:04:35.260
In supervised learning you're given
00:04:35.260 --> 00:04:40.100
pairs XY of some features and labels.
00:04:40.990 --> 00:04:42.430
And then you're trying to solve for
00:04:42.430 --> 00:04:45.570
parameters that minimizes your loss,
00:04:45.570 --> 00:04:49.909
and your loss is a is like A is a
00:04:49.910 --> 00:04:51.628
objective function that you're trying
00:04:51.628 --> 00:04:54.130
to reduce, and it usually has two
00:04:54.130 --> 00:04:54.860
components.
00:04:54.860 --> 00:04:56.550
1 component is that you want your
00:04:56.550 --> 00:04:59.140
predictions on the training data to be
00:04:59.140 --> 00:05:00.490
as good as possible.
00:05:00.490 --> 00:05:03.066
For example, you might say that you
00:05:03.066 --> 00:05:05.525
want to maximize the probability of
00:05:05.525 --> 00:05:07.210
your labels given your features.
00:05:07.840 --> 00:05:10.930
Or, equivalently, you want to minimize
00:05:10.930 --> 00:05:12.795
the negative sum of log likelihood of
00:05:12.795 --> 00:05:14.450
your labels given your features.
00:05:14.450 --> 00:05:16.762
This is the same as maximizing the
00:05:16.762 --> 00:05:17.650
likelihood of the labels.
00:05:18.280 --> 00:05:22.360
But we often want to minimize things,
00:05:22.360 --> 00:05:24.679
so negative log is.
00:05:24.680 --> 00:05:26.581
Minimizing the negative log is the same
00:05:26.581 --> 00:05:30.056
as maximizing the log and taking the
00:05:30.056 --> 00:05:30.369
log.
00:05:30.369 --> 00:05:33.500
The Max of the log is the same as the
00:05:33.500 --> 00:05:35.040
Max of the value.
00:05:35.850 --> 00:05:37.510
And this form tends to be easier to
00:05:37.510 --> 00:05:38.030
optimize.
00:05:40.730 --> 00:05:41.730
The second term.
00:05:41.730 --> 00:05:43.720
So we want to maximize the likelihood
00:05:43.720 --> 00:05:45.590
of the labels given the data, but we
00:05:45.590 --> 00:05:49.000
also want to have some likely.
00:05:49.000 --> 00:05:51.750
We often want to impose some kinds of
00:05:51.750 --> 00:05:53.450
constraints or some kinds of
00:05:53.450 --> 00:05:56.020
preferences for the parameters of our
00:05:56.020 --> 00:05:56.450
model.
00:05:57.210 --> 00:05:58.240
So.
00:05:58.430 --> 00:05:59.010
And.
00:06:00.730 --> 00:06:02.549
So a common thing is that we want to
00:06:02.550 --> 00:06:04.449
say that the sum of the parameters we
00:06:04.449 --> 00:06:06.119
want to minimize the sum of the
00:06:06.120 --> 00:06:07.465
parameter squared, or we want to
00:06:07.465 --> 00:06:09.148
minimize the sum of the absolute values
00:06:09.148 --> 00:06:10.282
of the parameters.
00:06:10.282 --> 00:06:11.815
So this is called regularization.
00:06:11.815 --> 00:06:13.565
Or if you have a probabilistic model,
00:06:13.565 --> 00:06:16.490
that might be in the form of a prior on
00:06:16.490 --> 00:06:19.200
the statistics that you're estimating.
00:06:20.910 --> 00:06:22.850
So the regularization and priors
00:06:22.850 --> 00:06:24.720
indicate some kind of preference for a
00:06:24.720 --> 00:06:26.110
particular solutions.
00:06:26.940 --> 00:06:28.700
And they tend to improve
00:06:28.700 --> 00:06:29.330
generalization.
00:06:29.330 --> 00:06:31.700
And in some cases they're necessary to
00:06:31.700 --> 00:06:33.520
obtain a unique solution.
00:06:33.520 --> 00:06:35.430
Like there might be many linear models
00:06:35.430 --> 00:06:37.510
that can separate your one class from
00:06:37.510 --> 00:06:40.380
another, and without regularization you
00:06:40.380 --> 00:06:41.820
have no way of choosing among those
00:06:41.820 --> 00:06:42.510
different models.
00:06:42.510 --> 00:06:45.660
The regularization specifies a
00:06:45.660 --> 00:06:46.720
particular solution.
00:06:48.250 --> 00:06:50.690
And this is it's more important the
00:06:50.690 --> 00:06:52.040
less data you have.
00:06:52.950 --> 00:06:55.450
Or the more features or larger your
00:06:55.450 --> 00:06:56.060
problem is.
00:07:00.900 --> 00:07:03.240
Once we've once we've trained a model,
00:07:03.240 --> 00:07:05.020
then we want to do prediction using
00:07:05.020 --> 00:07:06.033
that model.
00:07:06.033 --> 00:07:08.300
So in prediction we're given some new
00:07:08.300 --> 00:07:09.200
set of features.
00:07:09.860 --> 00:07:10.980
It will be the same.
00:07:10.980 --> 00:07:14.216
So in training we might have seen 500
00:07:14.216 --> 00:07:16.870
examples, and for each of those
00:07:16.870 --> 00:07:19.191
examples 10 features and some label
00:07:19.191 --> 00:07:20.716
you're trying to predict.
00:07:20.716 --> 00:07:23.480
So in testing you'll have a set of
00:07:23.480 --> 00:07:25.860
testing examples, and each one will
00:07:25.860 --> 00:07:27.272
also have the same number of features.
00:07:27.272 --> 00:07:29.028
So it might have 10 features as well,
00:07:29.028 --> 00:07:30.771
and you're trying to predict the same
00:07:30.771 --> 00:07:31.019
label.
00:07:31.020 --> 00:07:32.810
But in testing you don't give the model
00:07:32.810 --> 00:07:34.265
your label, you're trying to output the
00:07:34.265 --> 00:07:34.510
label.
00:07:35.550 --> 00:07:37.990
So in testing, we're given some test
00:07:37.990 --> 00:07:42.687
sample with input features XT and if
00:07:42.687 --> 00:07:44.084
we're doing a regression, then we're
00:07:44.084 --> 00:07:45.474
trying to output yet directly.
00:07:45.474 --> 00:07:48.050
So we're trying to say, predict the
00:07:48.050 --> 00:07:49.830
stock price or temperature or something
00:07:49.830 --> 00:07:50.520
like that.
00:07:50.520 --> 00:07:52.334
If we're doing classification, we're
00:07:52.334 --> 00:07:54.210
trying to output the likelihood of a
00:07:54.210 --> 00:07:56.065
particular category or the most likely
00:07:56.065 --> 00:07:56.470
category.
00:08:03.280 --> 00:08:03.780
And.
00:08:04.810 --> 00:08:08.590
So then there's a so if we're trying to
00:08:08.590 --> 00:08:10.490
develop a machine learning algorithm.
00:08:11.240 --> 00:08:13.780
Then we go through this model
00:08:13.780 --> 00:08:15.213
evaluation process.
00:08:15.213 --> 00:08:18.660
So the first step is that we need to
00:08:18.660 --> 00:08:19.930
collect some data.
00:08:19.930 --> 00:08:21.848
So if we're creating a new problem,
00:08:21.848 --> 00:08:25.900
then we might need to capture capture
00:08:25.900 --> 00:08:28.940
images or record observations or
00:08:28.940 --> 00:08:30.950
download information from the Internet,
00:08:30.950 --> 00:08:31.930
or whatever.
00:08:31.930 --> 00:08:33.688
One way or another, you need to get
00:08:33.688 --> 00:08:34.092
some data.
00:08:34.092 --> 00:08:36.232
You need to get labels for that data.
00:08:36.232 --> 00:08:37.550
So it might include.
00:08:37.550 --> 00:08:38.970
You might need to do some manual
00:08:38.970 --> 00:08:39.494
annotation.
00:08:39.494 --> 00:08:41.590
You might need to.
00:08:41.650 --> 00:08:44.080
Crowd source or use platforms to get
00:08:44.080 --> 00:08:44.820
the labels.
00:08:44.820 --> 00:08:46.760
At the end of this you'll have a whole
00:08:46.760 --> 00:08:49.708
set of samples X&Y where X are the are
00:08:49.708 --> 00:08:51.355
the features that you want to use to
00:08:51.355 --> 00:08:53.070
make a prediction and why are the
00:08:53.070 --> 00:08:54.625
predictions that you want to make.
00:08:54.625 --> 00:08:57.442
And then you split that data into a
00:08:57.442 --> 00:09:00.190
training and validation and a test set
00:09:00.190 --> 00:09:01.630
where you're going to use the training
00:09:01.630 --> 00:09:03.175
set to optimize parameters, validation
00:09:03.175 --> 00:09:06.130
set to choose your best model and
00:09:06.130 --> 00:09:08.070
testing for your final evaluation and
00:09:08.070 --> 00:09:08.680
performance.
00:09:10.180 --> 00:09:12.330
So once you have the data, you might
00:09:12.330 --> 00:09:14.134
spend some time inspecting the features
00:09:14.134 --> 00:09:16.605
and trying to understand the problem a
00:09:16.605 --> 00:09:17.203
little bit better.
00:09:17.203 --> 00:09:19.570
Trying to look at do some little test
00:09:19.570 --> 00:09:23.960
to see how like baselines work and how
00:09:23.960 --> 00:09:27.320
certain features predict the label.
00:09:28.410 --> 00:09:29.600
And then you'll decide on some
00:09:29.600 --> 00:09:31.190
candidate models and parameters.
00:09:31.870 --> 00:09:34.610
Then for each candidate you would train
00:09:34.610 --> 00:09:36.970
the parameters using the train set.
00:09:37.720 --> 00:09:39.970
And you'll evaluate your trained model
00:09:39.970 --> 00:09:41.170
on the validation set.
00:09:41.910 --> 00:09:43.870
And then you choose the best model
00:09:43.870 --> 00:09:45.630
based on your validation performance.
00:09:46.470 --> 00:09:48.800
And then you evaluate it on the test
00:09:48.800 --> 00:09:49.040
set.
00:09:50.320 --> 00:09:54.160
And sometimes, very often you have like
00:09:54.160 --> 00:09:55.320
a tree and vowel test set.
00:09:55.320 --> 00:09:56.920
But an alternative is that you could do
00:09:56.920 --> 00:09:59.320
cross validation, which I'll show an
00:09:59.320 --> 00:10:02.000
example of, where you just split your
00:10:02.000 --> 00:10:05.305
whole set into 10 parts and each time
00:10:05.305 --> 00:10:07.423
you train on 9 parts and test on the
00:10:07.423 --> 00:10:08.130
10th part.
00:10:08.130 --> 00:10:09.430
That becomes.
00:10:09.430 --> 00:10:11.410
If you have like a very limited amount
00:10:11.410 --> 00:10:13.070
of data then that can help you make the
00:10:13.070 --> 00:10:14.360
best use of your limited data.
00:10:16.340 --> 00:10:18.040
So typically when you're evaluating the
00:10:18.040 --> 00:10:19.500
performance, you're going to measure
00:10:19.500 --> 00:10:21.370
like the error, the accuracy like root
00:10:21.370 --> 00:10:23.250
mean squared error or accuracy, or the
00:10:23.250 --> 00:10:24.810
amount of variance you can explain.
00:10:26.070 --> 00:10:27.130
Or you could be doing.
00:10:27.130 --> 00:10:28.555
If you're doing like a retrieval task,
00:10:28.555 --> 00:10:30.190
you might do precision recall.
00:10:30.190 --> 00:10:31.600
So there's a variety of metrics that
00:10:31.600 --> 00:10:32.510
depend on the problem.
00:10:36.890 --> 00:10:37.390
So.
00:10:38.160 --> 00:10:39.730
When we're trying to think about like
00:10:39.730 --> 00:10:41.530
these mill algorithms, there's actually
00:10:41.530 --> 00:10:43.400
a lot of different things that we
00:10:43.400 --> 00:10:44.170
should consider.
00:10:45.300 --> 00:10:48.187
One of them is like, what is the model?
00:10:48.187 --> 00:10:50.330
What kinds of things can it represent?
00:10:50.330 --> 00:10:52.139
For example, in a linear model and a
00:10:52.140 --> 00:10:55.350
classifier model, it means that all the
00:10:55.350 --> 00:10:57.382
data that's on one side of the
00:10:57.382 --> 00:10:58.893
hyperplane is going to be assigned to
00:10:58.893 --> 00:11:00.619
one class, and all the data on the
00:11:00.620 --> 00:11:02.066
other side of the hyperplane will be
00:11:02.066 --> 00:11:04.210
assigned to another class, where for
00:11:04.210 --> 00:11:06.610
nearest neighbor you can have much more
00:11:06.610 --> 00:11:08.150
flexible decision boundaries.
00:11:10.010 --> 00:11:11.270
You can also think about.
00:11:11.270 --> 00:11:13.440
Maybe the model implies that some kinds
00:11:13.440 --> 00:11:16.160
of functions are preferred over others.
00:11:18.810 --> 00:11:20.470
You think about like what is your
00:11:20.470 --> 00:11:21.187
objective function?
00:11:21.187 --> 00:11:22.870
So what is it that you're trying to
00:11:22.870 --> 00:11:25.100
minimize, and what kinds of like values
00:11:25.100 --> 00:11:26.060
does that imply?
00:11:26.060 --> 00:11:26.960
So do you prefer?
00:11:26.960 --> 00:11:27.890
Does it mean?
00:11:27.890 --> 00:11:29.840
Does your regularization, for example,
00:11:29.840 --> 00:11:32.620
mean that you prefer that you're using
00:11:32.620 --> 00:11:34.230
a few features or that you have low
00:11:34.230 --> 00:11:35.520
weight on a lot of features?
00:11:36.270 --> 00:11:39.126
Are you trying to minimize a likelihood
00:11:39.126 --> 00:11:42.190
or maximize the likelihood, or are you
00:11:42.190 --> 00:11:45.250
trying to just get high enough
00:11:45.250 --> 00:11:46.899
confidence on each example to get
00:11:46.900 --> 00:11:47.610
things correct?
00:11:49.430 --> 00:11:50.850
And it's important to note that the
00:11:50.850 --> 00:11:53.080
objective function often does not match
00:11:53.080 --> 00:11:54.290
your final evaluation.
00:11:54.290 --> 00:11:57.590
So nobody really trains a model to
00:11:57.590 --> 00:12:00.170
minimize the classification error, even
00:12:00.170 --> 00:12:01.960
though they often evaluate based on
00:12:01.960 --> 00:12:03.000
classification error.
00:12:03.940 --> 00:12:06.576
And there's two reasons for that.
00:12:06.576 --> 00:12:09.388
So one reason is that it's really hard
00:12:09.388 --> 00:12:11.550
to minimize classification error over
00:12:11.550 --> 00:12:13.510
training set, because a small change in
00:12:13.510 --> 00:12:15.000
parameters may not change your
00:12:15.000 --> 00:12:15.680
classification error.
00:12:15.680 --> 00:12:18.200
So it's hard to for an optimization
00:12:18.200 --> 00:12:19.850
algorithm to figure out how it should
00:12:19.850 --> 00:12:21.400
change to minimize that error.
00:12:22.430 --> 00:12:25.823
The second reason is that there might
00:12:25.823 --> 00:12:27.730
be many different models that can have
00:12:27.730 --> 00:12:29.620
similar classification error, the same
00:12:29.620 --> 00:12:31.980
classification error, and so you need
00:12:31.980 --> 00:12:33.560
some way of choosing among them.
00:12:33.560 --> 00:12:35.670
So many algorithms, many times the
00:12:35.670 --> 00:12:37.422
objective function will also say that
00:12:37.422 --> 00:12:39.160
you want to be very confident about
00:12:39.160 --> 00:12:41.274
your examples, not only that, you want
00:12:41.274 --> 00:12:42.010
to be correct.
00:12:45.380 --> 00:12:47.140
The third thing is that you would think
00:12:47.140 --> 00:12:50.070
about how you can optimize the model.
00:12:50.070 --> 00:12:51.610
So does it.
00:12:51.680 --> 00:12:56.200
For example for like logistic
00:12:56.200 --> 00:12:56.880
regression.
00:12:58.760 --> 00:13:01.480
You're able to reach a global optimum.
00:13:01.480 --> 00:13:04.220
It's a convex problem so that you're
00:13:04.220 --> 00:13:06.290
going to find the best solution, where
00:13:06.290 --> 00:13:08.020
for something a neural network it may
00:13:08.020 --> 00:13:09.742
not be possible to get the best
00:13:09.742 --> 00:13:11.000
solution, but you can usually get a
00:13:11.000 --> 00:13:11.860
pretty good solution.
00:13:12.680 --> 00:13:14.430
You also will think about like how long
00:13:14.430 --> 00:13:17.260
does it take to train and how does that
00:13:17.260 --> 00:13:18.709
depend on the number of examples and
00:13:18.709 --> 00:13:19.950
the number of features.
00:13:19.950 --> 00:13:22.010
So if you're later we'll talk about
00:13:22.010 --> 00:13:25.260
SVMS and Kernelized SVM is one of the
00:13:25.260 --> 00:13:27.560
problems, is that it's the training is
00:13:27.560 --> 00:13:29.761
quadratic in the number of examples, so
00:13:29.761 --> 00:13:32.600
it becomes a pretty expensive, at least
00:13:32.600 --> 00:13:34.976
according to the earlier optimization
00:13:34.976 --> 00:13:35.582
algorithms.
00:13:35.582 --> 00:13:38.120
So some algorithms can be used with a
00:13:38.120 --> 00:13:39.710
lot of examples, but some are just too
00:13:39.710 --> 00:13:40.370
expensive.
00:13:40.440 --> 00:13:40.880
Yeah.
00:13:43.520 --> 00:13:47.060
So the objective function is your, it's
00:13:47.060 --> 00:13:48.120
your loss essentially.
00:13:48.120 --> 00:13:50.470
So it usually has that data term where
00:13:50.470 --> 00:13:51.540
you're trying to maximize the
00:13:51.540 --> 00:13:52.910
likelihood of the data or the labels
00:13:52.910 --> 00:13:53.960
given the data.
00:13:53.960 --> 00:13:56.075
And it has some regularization term
00:13:56.075 --> 00:13:58.130
that says that you prefer some models
00:13:58.130 --> 00:13:58.670
over others.
00:14:05.090 --> 00:14:07.890
So yeah, feel free to please do ask as
00:14:07.890 --> 00:14:11.140
many questions as pop into your mind.
00:14:11.140 --> 00:14:13.010
I'm happy to answer them and I want to
00:14:13.010 --> 00:14:14.992
make sure, hopefully at the end of this
00:14:14.992 --> 00:14:17.670
lecture, or if it's or if you like
00:14:17.670 --> 00:14:18.630
further review the lecture.
00:14:18.630 --> 00:14:20.345
Again, I hope that all of this stuff is
00:14:20.345 --> 00:14:22.340
like really clear, and if it's not,
00:14:22.340 --> 00:14:26.847
just don't feel don't be afraid to ask
00:14:26.847 --> 00:14:28.680
questions in office hours or after
00:14:28.680 --> 00:14:29.660
class or whatever.
00:14:31.920 --> 00:14:34.065
So then finally, how does the
00:14:34.065 --> 00:14:34.670
prediction work?
00:14:34.670 --> 00:14:36.340
So then you want to think about like
00:14:36.340 --> 00:14:37.740
can I make a prediction really quickly?
00:14:37.740 --> 00:14:39.730
So like for a nearest neighbor it's not
00:14:39.730 --> 00:14:41.579
necessarily so quick, but for the
00:14:41.580 --> 00:14:43.050
linear models it's pretty fast.
00:14:44.750 --> 00:14:46.580
Can I find the most likely prediction
00:14:46.580 --> 00:14:48.260
according to my model?
00:14:48.260 --> 00:14:50.390
So sometimes even after you've
00:14:50.390 --> 00:14:53.790
optimized your model, you don't have a
00:14:53.790 --> 00:14:55.530
guarantee that you can generate the
00:14:55.530 --> 00:14:57.410
best solution for a new sample.
00:14:57.410 --> 00:14:59.930
So for example with these image
00:14:59.930 --> 00:15:02.090
generation algorithms even though.
00:15:02.890 --> 00:15:05.060
Even after you optimize your model
00:15:05.060 --> 00:15:08.150
given some phrase, you're not
00:15:08.150 --> 00:15:09.720
necessarily going to generate the most
00:15:09.720 --> 00:15:11.630
likely image given that phrase.
00:15:11.630 --> 00:15:13.710
You'll just generate like an image that
00:15:13.710 --> 00:15:16.199
is like consistent with the phrase
00:15:16.200 --> 00:15:18.010
according to some scoring function.
00:15:18.010 --> 00:15:20.810
So not all models can even be perfectly
00:15:20.810 --> 00:15:22.040
optimized for prediction.
00:15:23.100 --> 00:15:25.110
And then finally, does my algorithm
00:15:25.110 --> 00:15:27.180
output confidence as well as
00:15:27.180 --> 00:15:27.710
prediction?
00:15:27.710 --> 00:15:30.770
Usually it's helpful if your model not
00:15:30.770 --> 00:15:32.193
only gives you an answer, but also
00:15:32.193 --> 00:15:33.930
gives you a confidence in how to write
00:15:33.930 --> 00:15:34.790
that answer is.
00:15:35.420 --> 00:15:37.580
And it's nice if that confidence is
00:15:37.580 --> 00:15:38.030
accurate.
00:15:39.240 --> 00:15:41.580
Meaning that if it says that you've got
00:15:41.580 --> 00:15:44.000
like a 99% chance of being correct,
00:15:44.000 --> 00:15:46.250
then hopefully 99 out of 100 times
00:15:46.250 --> 00:15:48.640
you'll be correct in that situation.
00:15:55.440 --> 00:15:57.234
So we looked at.
00:15:57.234 --> 00:15:59.300
We looked at several different
00:15:59.300 --> 00:16:00.870
classification algorithms.
00:16:01.560 --> 00:16:04.440
And so here they're all compared
00:16:04.440 --> 00:16:05.890
side-by-side according to some
00:16:05.890 --> 00:16:06.290
criteria.
00:16:06.290 --> 00:16:08.130
So we can think about like what type of
00:16:08.130 --> 00:16:10.290
algorithm it is it a nearest neighbor
00:16:10.290 --> 00:16:12.480
is instance based, and that the
00:16:12.480 --> 00:16:14.120
parameters are the instances
00:16:14.120 --> 00:16:14.740
themselves.
00:16:14.740 --> 00:16:17.870
There's additional like linear model or
00:16:17.870 --> 00:16:19.450
something that's parametric that you're
00:16:19.450 --> 00:16:20.590
trying to fit to your data.
00:16:22.150 --> 00:16:24.170
Naive Bayes is probabilistic is
00:16:24.170 --> 00:16:26.060
logistic regression, but.
00:16:26.910 --> 00:16:29.090
Naive Bayes, you're maximizing the
00:16:29.090 --> 00:16:31.210
likelihood of your features given the
00:16:31.210 --> 00:16:33.020
data or your features, and I mean
00:16:33.020 --> 00:16:34.270
sorry, you're maximizing likelihood of
00:16:34.270 --> 00:16:35.720
your features and the label.
00:16:36.600 --> 00:16:37.230
00:16:38.790 --> 00:16:40.800
Under the assumption that your features
00:16:40.800 --> 00:16:42.610
are independent given the label.
00:16:43.450 --> 00:16:45.450
Where in logistic regression you're
00:16:45.450 --> 00:16:47.695
directly maximizing the likelihood of
00:16:47.695 --> 00:16:48.970
the label given the data.
00:16:51.820 --> 00:16:53.880
They both often end up being linear
00:16:53.880 --> 00:16:55.750
models, but you're modeling different
00:16:55.750 --> 00:16:57.659
things in these two in these two
00:16:57.660 --> 00:16:58.170
settings.
00:16:58.790 --> 00:17:01.880
And in logistic regression, the model
00:17:01.880 --> 00:17:04.460
the linear part, so it's, I just wrote
00:17:04.460 --> 00:17:05.890
logistic regression, but often we're
00:17:05.890 --> 00:17:07.176
doing linear logistic regression.
00:17:07.176 --> 00:17:09.490
The linear part is that we're seeing
00:17:09.490 --> 00:17:11.993
that this logic function is linear.
00:17:11.993 --> 00:17:15.896
The log ratio of the probability of the
00:17:15.896 --> 00:17:19.830
of label equals one given the features
00:17:19.830 --> 00:17:21.460
over probability of label equals zero
00:17:21.460 --> 00:17:22.319
given the features.
00:17:22.319 --> 00:17:24.323
That thing is the linear thing that
00:17:24.323 --> 00:17:24.970
we're fitting.
00:17:27.290 --> 00:17:28.700
And then we talked about decision
00:17:28.700 --> 00:17:29.350
trees.
00:17:29.350 --> 00:17:31.706
I would also say that's a kind of a
00:17:31.706 --> 00:17:33.040
probabilistic function in the sense
00:17:33.040 --> 00:17:35.555
that we're choosing our splits to
00:17:35.555 --> 00:17:38.700
maximize the mutual information or to,
00:17:38.700 --> 00:17:41.200
sorry, to maximize the information gain
00:17:41.200 --> 00:17:44.870
to minimize the conditional entropy.
00:17:44.870 --> 00:17:47.780
And that's like a probabilistic basis
00:17:47.780 --> 00:17:49.400
for the optimization.
00:17:50.080 --> 00:17:51.810
And then at the end of the prediction,
00:17:51.810 --> 00:17:53.560
you would typically be estimating the
00:17:53.560 --> 00:17:55.330
probability of each label given the
00:17:55.330 --> 00:17:57.170
data that has fallen into some leaf
00:17:57.170 --> 00:17:57.430
node.
00:17:59.490 --> 00:18:01.024
But that has quite different rules than
00:18:01.024 --> 00:18:01.460
the other.
00:18:01.460 --> 00:18:03.260
So nearest neighbor is just going to be
00:18:03.260 --> 00:18:05.189
like finding the sample that has the
00:18:05.190 --> 00:18:06.750
closest distance.
00:18:06.750 --> 00:18:08.422
Naive Bayes and logistic regression
00:18:08.422 --> 00:18:11.363
will be these probability functions
00:18:11.363 --> 00:18:13.540
that will tend to give you like linear
00:18:13.540 --> 00:18:14.485
classifiers.
00:18:14.485 --> 00:18:17.480
And Decision Tree has these conjunctive
00:18:17.480 --> 00:18:19.840
rules that you say if this feature is
00:18:19.840 --> 00:18:22.249
greater than this value then you go
00:18:22.249 --> 00:18:22.615
this way.
00:18:22.615 --> 00:18:23.955
And then if this other thing happens
00:18:23.955 --> 00:18:26.090
then you go another way and then at the
00:18:26.090 --> 00:18:29.350
end you can express that as a series of
00:18:29.350 --> 00:18:29.700
rules.
00:18:29.750 --> 00:18:31.425
Where you have a bunch of and
00:18:31.425 --> 00:18:32.850
conditions, and if all of those
00:18:32.850 --> 00:18:34.220
conditions are met, then you make a
00:18:34.220 --> 00:18:35.290
particular prediction.
00:18:38.370 --> 00:18:40.150
So these algorithms have different
00:18:40.150 --> 00:18:42.480
strengths, like nearest neighbor has
00:18:42.480 --> 00:18:45.547
low bias, so that means that you can
00:18:45.547 --> 00:18:47.340
almost always get perfect training
00:18:47.340 --> 00:18:47.970
accuracy.
00:18:47.970 --> 00:18:49.706
You can fit like almost anything with
00:18:49.706 --> 00:18:50.279
nearest neighbor.
00:18:52.310 --> 00:18:54.725
On the other hand, I guess I didn't put
00:18:54.725 --> 00:18:56.640
it here, but limitation is that it has
00:18:56.640 --> 00:18:57.300
high variance.
00:18:58.000 --> 00:18:59.650
You might get very different prediction
00:18:59.650 --> 00:19:01.590
functions if you resample your data.
00:19:03.390 --> 00:19:05.150
It has no training time.
00:19:06.230 --> 00:19:08.200
It's very widely applicable and it's
00:19:08.200 --> 00:19:08.900
very simple.
00:19:09.690 --> 00:19:12.110
Another limitation is that it can take
00:19:12.110 --> 00:19:13.780
a long time to do inference, but if you
00:19:13.780 --> 00:19:15.642
use approximate nearest neighbor
00:19:15.642 --> 00:19:17.790
inference, which we'll talk about
00:19:17.790 --> 00:19:21.230
later, then it can be like relatively
00:19:21.230 --> 00:19:21.608
fast.
00:19:21.608 --> 00:19:23.881
You can do approximate nearest neighbor
00:19:23.881 --> 00:19:26.600
in log N time, where N is the number of
00:19:26.600 --> 00:19:29.310
training samples, where so far we're
00:19:29.310 --> 00:19:31.470
just doing brute force, which is linear
00:19:31.470 --> 00:19:32.460
in the number of samples.
00:19:34.620 --> 00:19:35.770
Naive bayes.
00:19:35.770 --> 00:19:37.980
The strengths are that you can estimate
00:19:37.980 --> 00:19:39.950
these parameters reasonably well from
00:19:39.950 --> 00:19:40.680
limited data.
00:19:41.690 --> 00:19:43.000
It's also pretty simple.
00:19:43.000 --> 00:19:45.380
It's fast to train, and the downside is
00:19:45.380 --> 00:19:48.030
that as limited modeling power, so even
00:19:48.030 --> 00:19:49.876
on the training set you often won't get
00:19:49.876 --> 00:19:52.049
0 error or even close to 0 error.
00:19:53.520 --> 00:19:55.290
Logistic regression is really powerful
00:19:55.290 --> 00:19:57.250
in high dimensions, so remember that
00:19:57.250 --> 00:19:59.050
even though it's a linear classifier,
00:19:59.050 --> 00:20:01.400
which feels like it can't do much in
00:20:01.400 --> 00:20:04.830
terms of separation in high dimensions,
00:20:04.830 --> 00:20:05.530
you can.
00:20:05.530 --> 00:20:07.330
These classifiers are actually very
00:20:07.330 --> 00:20:07.850
powerful.
00:20:08.510 --> 00:20:10.710
If you have 1000 dimensional feature.
00:20:11.330 --> 00:20:13.930
And you have 1000 data points, then you
00:20:13.930 --> 00:20:16.094
can assign those data points arbitrary
00:20:16.094 --> 00:20:18.210
labels, arbitrary binary labels, and
00:20:18.210 --> 00:20:19.590
still get a perfect classifier.
00:20:19.590 --> 00:20:21.770
You're guaranteed a perfect classifier
00:20:21.770 --> 00:20:23.050
in terms of the training data.
00:20:23.860 --> 00:20:26.740
Now, that power power is always a
00:20:26.740 --> 00:20:27.750
double edged sword.
00:20:27.750 --> 00:20:29.740
You, if you have a powerful classifier,
00:20:29.740 --> 00:20:32.040
means you can fit your training data
00:20:32.040 --> 00:20:34.140
really well, but it also means that
00:20:34.140 --> 00:20:35.850
you're more susceptible to overfitting
00:20:35.850 --> 00:20:37.510
your training data, which means that
00:20:37.510 --> 00:20:38.510
you perform well.
00:20:39.460 --> 00:20:41.160
And the training data, but your test
00:20:41.160 --> 00:20:43.170
performance is not so good, you get
00:20:43.170 --> 00:20:43.940
higher test error.
00:20:45.780 --> 00:20:47.830
It's also widely applicable.
00:20:47.830 --> 00:20:50.480
It produces good confidence estimates,
00:20:50.480 --> 00:20:52.130
so that can be helpful if you want to
00:20:52.130 --> 00:20:54.170
know whether the prediction is correct.
00:20:54.780 --> 00:20:56.640
And it gives you fast prediction
00:20:56.640 --> 00:20:57.840
because it's the linear model.
00:20:59.470 --> 00:21:01.470
Similar to nearest neighbor has a
00:21:01.470 --> 00:21:03.380
limitation that it relies on good input
00:21:03.380 --> 00:21:04.330
features.
00:21:04.330 --> 00:21:05.730
So nearest neighbor if you have a
00:21:05.730 --> 00:21:06.160
simple.
00:21:07.240 --> 00:21:10.040
If you have a simple distance function
00:21:10.040 --> 00:21:13.660
like Euclidian distance, that assumes
00:21:13.660 --> 00:21:15.665
that all your features are scaled so
00:21:15.665 --> 00:21:17.110
that there are like comparable scales
00:21:17.110 --> 00:21:18.930
to each other, and that they're all
00:21:18.930 --> 00:21:19.540
predictive.
00:21:20.400 --> 00:21:22.310
Nearest logistic regression doesn't
00:21:22.310 --> 00:21:23.970
make assumptions that strong.
00:21:23.970 --> 00:21:25.799
It can kind of choose which features to
00:21:25.800 --> 00:21:27.420
use and it can rescale them
00:21:27.420 --> 00:21:29.790
essentially, but it does.
00:21:29.790 --> 00:21:33.230
But it's not able to model like joint
00:21:33.230 --> 00:21:35.425
combinations of features, so the
00:21:35.425 --> 00:21:37.360
features should be individually useful.
00:21:39.270 --> 00:21:41.340
And then finally, decision trees are
00:21:41.340 --> 00:21:42.930
good because they can provide an
00:21:42.930 --> 00:21:44.600
explainable decision function.
00:21:44.600 --> 00:21:47.040
You get these nice rules that are easy
00:21:47.040 --> 00:21:47.750
to communicate.
00:21:48.360 --> 00:21:49.740
It's also widely applicable.
00:21:49.740 --> 00:21:51.400
You can use that on continuous discrete
00:21:51.400 --> 00:21:52.040
data.
00:21:52.040 --> 00:21:54.162
You don't need to scale the features.
00:21:54.162 --> 00:21:55.740
It's like it doesn't really matter if
00:21:55.740 --> 00:21:57.930
you multiply the features by 10, it
00:21:57.930 --> 00:21:59.230
just means that you'd be choosing a
00:21:59.230 --> 00:22:00.790
threshold that's 10 times bigger.
00:22:01.820 --> 00:22:03.510
And you can deal with a mix of discrete
00:22:03.510 --> 00:22:05.720
and continuous variables.
00:22:05.720 --> 00:22:07.380
The downside is that.
00:22:08.330 --> 00:22:11.780
One tree by itself either tends to
00:22:11.780 --> 00:22:14.170
generalize poorly, meaning like you
00:22:14.170 --> 00:22:15.870
train a full tree and you do perfect
00:22:15.870 --> 00:22:18.140
training, but you get bad test error.
00:22:18.770 --> 00:22:20.240
Or you tend to underfit the data.
00:22:20.240 --> 00:22:21.910
If you train a short tree then you
00:22:21.910 --> 00:22:23.510
don't get very good training or test
00:22:23.510 --> 00:22:23.770
error.
00:22:24.650 --> 00:22:26.920
And so a single tree by itself is not
00:22:26.920 --> 00:22:28.160
usually the best predictor.
00:22:31.530 --> 00:22:34.085
So there's just like you can also think
00:22:34.085 --> 00:22:35.530
about these methods, I won't talk
00:22:35.530 --> 00:22:37.366
through this whole slide, but you can
00:22:37.366 --> 00:22:39.290
also think about the methods in terms
00:22:39.290 --> 00:22:42.130
of like the learning objectives, the
00:22:42.130 --> 00:22:44.556
training, like how you optimize those
00:22:44.556 --> 00:22:46.350
learning objectives and then the
00:22:46.350 --> 00:22:47.840
inference, how you make your final
00:22:47.840 --> 00:22:48.430
prediction.
00:22:49.040 --> 00:22:52.460
And so here I also included linear
00:22:52.460 --> 00:22:54.870
SVMS, which we'll talk about next week,
00:22:54.870 --> 00:22:57.590
but you can see for example that.
00:22:59.260 --> 00:23:01.730
That these in terms of inference,
00:23:01.730 --> 00:23:04.200
linear SVM, logistic regression, Naive
00:23:04.200 --> 00:23:06.790
Bayes are all linear models, at least
00:23:06.790 --> 00:23:08.230
in the case where you're dealing with
00:23:08.230 --> 00:23:11.190
discrete variables or Gaussians for 9
00:23:11.190 --> 00:23:11.630
days.
00:23:11.630 --> 00:23:13.948
But they have different ways, they have
00:23:13.948 --> 00:23:15.695
different learning objectives and then
00:23:15.695 --> 00:23:17.000
different ways of doing the training.
00:23:22.330 --> 00:23:24.790
And then question go ahead.
00:23:36.030 --> 00:23:37.450
Yeah.
00:23:37.450 --> 00:23:39.810
Thank you for the clarification, so.
00:23:40.710 --> 00:23:42.850
So what I mean by that it doesn't
00:23:42.850 --> 00:23:46.110
require feature scaling is that if you
00:23:46.110 --> 00:23:47.909
could have one feature that ranges from
00:23:47.910 --> 00:23:50.495
like zero to 1000 and another feature
00:23:50.495 --> 00:23:52.160
that ranges from zero to 1.
00:23:52.960 --> 00:23:56.090
And decision trees are perfectly fine
00:23:56.090 --> 00:23:57.770
with that, because it can like freely
00:23:57.770 --> 00:23:59.390
choose the threshold and stuff.
00:23:59.390 --> 00:24:01.450
And if you multiply 1 feature value by
00:24:01.450 --> 00:24:03.700
50, it doesn't really change the
00:24:03.700 --> 00:24:05.643
function, it can still choose like
00:24:05.643 --> 00:24:07.300
threshold that's 50 times larger.
00:24:08.050 --> 00:24:10.220
Where nearest neighbor, for example, if
00:24:10.220 --> 00:24:13.084
one feature ranges from zero to 1001
00:24:13.084 --> 00:24:15.880
ranges from zero to 1, then it's not
00:24:15.880 --> 00:24:17.673
going to care at all about the zero to
00:24:17.673 --> 00:24:19.270
1 feature because like that difference
00:24:19.270 --> 00:24:21.790
of like 200 on the scale of zero to
00:24:21.790 --> 00:24:23.738
1000 is going to overwhelm completely a
00:24:23.738 --> 00:24:26.290
difference of 1 on the 0 to one
00:24:26.290 --> 00:24:26.609
feature.
00:24:35.130 --> 00:24:36.275
Right, it doesn't.
00:24:36.275 --> 00:24:37.910
It's not influenced.
00:24:37.910 --> 00:24:40.040
I guess it's not influenced by the
00:24:40.040 --> 00:24:41.370
variance of the features, yeah.
00:24:46.320 --> 00:24:49.130
So I don't need to read talk through
00:24:49.130 --> 00:24:51.260
all of this because even for
00:24:51.260 --> 00:24:53.480
aggression, most of these algorithms
00:24:53.480 --> 00:24:55.219
are the same and they have the same
00:24:55.220 --> 00:24:56.710
strengths and the same weaknesses.
00:24:57.500 --> 00:24:59.630
The only difference between regression
00:24:59.630 --> 00:25:01.310
and classification is that you tend to
00:25:01.310 --> 00:25:03.235
have a different loss function where
00:25:03.235 --> 00:25:04.820
you because you're trying to predict a
00:25:04.820 --> 00:25:06.790
continuous value instead of predicting
00:25:06.790 --> 00:25:09.590
a likelihood of a categorical value, or
00:25:09.590 --> 00:25:11.240
trying to just output the categorical
00:25:11.240 --> 00:25:12.000
value directly.
00:25:14.330 --> 00:25:17.450
Linear regression though is A1 new
00:25:17.450 --> 00:25:18.290
algorithm here.
00:25:18.980 --> 00:25:21.923
So in linear regression, you're trying
00:25:21.923 --> 00:25:24.585
to fit the data, so you're not trying
00:25:24.585 --> 00:25:24.940
to.
00:25:26.480 --> 00:25:28.396
Fit like a probability model like
00:25:28.396 --> 00:25:29.590
linear logistic regression.
00:25:29.590 --> 00:25:31.860
You're just trying to directly fit the
00:25:31.860 --> 00:25:33.680
prediction given the data, and so you
00:25:33.680 --> 00:25:35.575
have like a linear function like W
00:25:35.575 --> 00:25:37.960
transpose X or West transpose X + B.
00:25:37.960 --> 00:25:41.120
That should ideally output output Y
00:25:41.120 --> 00:25:41.710
directly.
00:25:43.830 --> 00:25:45.670
Similar to linear to logistic
00:25:45.670 --> 00:25:47.030
regression, though it's powerful and
00:25:47.030 --> 00:25:48.220
high dimensions, it's widely
00:25:48.220 --> 00:25:48.820
applicable.
00:25:48.820 --> 00:25:50.650
You get fast prediction.
00:25:50.650 --> 00:25:52.770
Also, it can be useful to interpret the
00:25:52.770 --> 00:25:54.300
coefficients to say like what the
00:25:54.300 --> 00:25:56.040
correlations are of the features with
00:25:56.040 --> 00:25:58.110
your prediction, or to see which
00:25:58.110 --> 00:25:59.900
features are more predictive than
00:25:59.900 --> 00:26:00.300
others.
00:26:01.410 --> 00:26:03.440
And similar to logistic regression, it
00:26:03.440 --> 00:26:06.140
relies to some extent on good features.
00:26:06.140 --> 00:26:07.720
In fact, I would say even more.
00:26:08.320 --> 00:26:12.220
Because this is assuming that Y is
00:26:12.220 --> 00:26:15.040
going to be a linear function of X and
00:26:15.040 --> 00:26:17.130
West, which is in a way a stronger
00:26:17.130 --> 00:26:18.140
assumption than that.
00:26:18.140 --> 00:26:20.670
Like a binary classification will be a
00:26:20.670 --> 00:26:21.870
linear function of the features.
00:26:23.360 --> 00:26:24.940
So you often have to do some kind of
00:26:24.940 --> 00:26:26.950
feature transformations to make it work
00:26:26.950 --> 00:26:27.220
well.
00:26:28.520 --> 00:26:28.960
Question.
00:26:40.800 --> 00:26:43.402
So naive bayes.
00:26:43.402 --> 00:26:46.295
The example I gave was a semi semi
00:26:46.295 --> 00:26:48.850
Naive Bayes algorithm for classifying
00:26:48.850 --> 00:26:50.650
faces and cars.
00:26:50.650 --> 00:26:52.618
So there they took groups of features
00:26:52.618 --> 00:26:54.190
and modeled the probabilities of small
00:26:54.190 --> 00:26:55.720
groups of features and then took the
00:26:55.720 --> 00:26:57.090
product of those to give you your
00:26:57.090 --> 00:26:58.190
probabilistic model.
00:26:58.190 --> 00:27:01.770
I also would use like Naive Bayes if
00:27:01.770 --> 00:27:03.719
I'm trying to do like color like
00:27:03.720 --> 00:27:05.600
segmentation based on color and I need
00:27:05.600 --> 00:27:08.000
to estimate the probability of color
00:27:08.000 --> 00:27:09.490
given that it's in one region versus
00:27:09.490 --> 00:27:11.470
another, I might assume that.
00:27:11.530 --> 00:27:15.320
By that, my color features like the hue
00:27:15.320 --> 00:27:17.920
versus intensity for example, are
00:27:17.920 --> 00:27:19.380
independent given the region that it
00:27:19.380 --> 00:27:22.260
came from and so use that as part of my
00:27:22.260 --> 00:27:23.760
probabilistic model for doing the
00:27:23.760 --> 00:27:24.670
segmentation.
00:27:25.880 --> 00:27:30.940
Logistic regression you would like any
00:27:30.940 --> 00:27:32.610
neural network is doing logistic
00:27:32.610 --> 00:27:35.807
regression in the last layer.
00:27:35.807 --> 00:27:38.703
So most things are using logistic
00:27:38.703 --> 00:27:40.770
regression now as part of it.
00:27:40.770 --> 00:27:42.775
So you can view like the early layers
00:27:42.775 --> 00:27:44.674
as feature learning and the last layer
00:27:44.674 --> 00:27:45.519
is logistic regression.
00:27:46.490 --> 00:27:49.250
And then decision trees are.
00:27:50.660 --> 00:27:52.200
We'll see an example.
00:27:52.200 --> 00:27:53.670
It's used in the example I'm going to
00:27:53.670 --> 00:27:55.723
give, but like medical analysis is a is
00:27:55.723 --> 00:27:57.680
a good one because you often want some
00:27:57.680 --> 00:28:00.631
interpretable function as well as some
00:28:00.631 --> 00:28:01.620
good prediction.
00:28:03.820 --> 00:28:04.090
Yep.
00:28:09.200 --> 00:28:11.450
All right, so one of the one of the key
00:28:11.450 --> 00:28:15.360
concepts is like how performance varies
00:28:15.360 --> 00:28:17.230
with the number of training samples.
00:28:17.230 --> 00:28:20.080
So as you get more training data, you
00:28:20.080 --> 00:28:21.670
should be able to fit a more accurate
00:28:21.670 --> 00:28:22.120
model.
00:28:23.310 --> 00:28:25.600
And so you would expect that your test
00:28:25.600 --> 00:28:27.746
error should decrease as you get more
00:28:27.746 --> 00:28:29.760
training samples, because if you have
00:28:29.760 --> 00:28:33.640
only like 1 training sample, then you
00:28:33.640 --> 00:28:34.700
don't know if that's like really
00:28:34.700 --> 00:28:36.420
representative, if it's covering all
00:28:36.420 --> 00:28:37.195
the different cases.
00:28:37.195 --> 00:28:39.263
As you get more and more training
00:28:39.263 --> 00:28:41.020
samples, you can fit more complex
00:28:41.020 --> 00:28:43.858
models and you can be more assured that
00:28:43.858 --> 00:28:46.110
the training samples that you've seen
00:28:46.110 --> 00:28:47.850
fully represent the distribution that
00:28:47.850 --> 00:28:48.710
you'll see in testing.
00:28:50.040 --> 00:28:52.040
But as you get more training, it
00:28:52.040 --> 00:28:53.700
becomes harder to fit the training
00:28:53.700 --> 00:28:54.060
data.
00:28:54.920 --> 00:28:57.655
So maybe a linear model can perfectly
00:28:57.655 --> 00:29:00.340
classify like 500 examples, but it
00:29:00.340 --> 00:29:02.350
can't perfectly classify 500 million
00:29:02.350 --> 00:29:04.900
examples, even if they're even in the
00:29:04.900 --> 00:29:05.430
training set.
00:29:07.110 --> 00:29:10.420
As you get more data, these will test
00:29:10.420 --> 00:29:12.630
and the training error will converge.
00:29:13.380 --> 00:29:15.100
And if they're coming from exactly the
00:29:15.100 --> 00:29:16.540
same distribution, then they'll
00:29:16.540 --> 00:29:18.500
converge to exactly the same value.
00:29:19.680 --> 00:29:21.030
Only if they come from different
00:29:21.030 --> 00:29:22.790
distributions would you possibly have a
00:29:22.790 --> 00:29:24.250
gap if you have infinite training
00:29:24.250 --> 00:29:24.720
samples.
00:29:25.330 --> 00:29:27.133
So we have these concepts of the test
00:29:27.133 --> 00:29:27.411
error.
00:29:27.411 --> 00:29:29.253
So that's the error on some samples
00:29:29.253 --> 00:29:31.420
that are not used for training that are
00:29:31.420 --> 00:29:34.360
randomly sampled from your distribution
00:29:34.360 --> 00:29:35.020
of data.
00:29:35.020 --> 00:29:38.744
The training error is the error on your
00:29:38.744 --> 00:29:41.240
training set that is used to optimize
00:29:41.240 --> 00:29:43.458
your model, and the generalization
00:29:43.458 --> 00:29:46.803
error is the gap between the test and
00:29:46.803 --> 00:29:49.237
the training error, so that the
00:29:49.237 --> 00:29:51.672
generalization error is your error due
00:29:51.672 --> 00:29:55.386
to due to like an imperfect model due
00:29:55.386 --> 00:29:55.679
to.
00:29:55.750 --> 00:29:57.280
To limited training samples.
00:30:04.950 --> 00:30:05.650
Question.
00:30:07.940 --> 00:30:09.675
So there's test error.
00:30:09.675 --> 00:30:12.620
So that's the I'll start with training.
00:30:12.620 --> 00:30:14.070
OK, so first there's training error.
00:30:14.810 --> 00:30:17.610
So training error is you fit, you fit a
00:30:17.610 --> 00:30:19.010
model on a training set, and then
00:30:19.010 --> 00:30:20.540
you're evaluating the error on the same
00:30:20.540 --> 00:30:21.230
training set.
00:30:22.490 --> 00:30:24.620
So if your model is really powerful,
00:30:24.620 --> 00:30:27.282
that training error might be 0, But if
00:30:27.282 --> 00:30:29.220
it's if it's more limited, like Naive
00:30:29.220 --> 00:30:32.090
Bayes, you'll often have nonzero error.
00:30:32.950 --> 00:30:35.652
And you since your loss is, since you
00:30:35.652 --> 00:30:36.384
have some.
00:30:36.384 --> 00:30:38.580
If you're optimizing a loss like the
00:30:38.580 --> 00:30:41.160
probability, then there's always room
00:30:41.160 --> 00:30:42.870
to improve that loss, so you'll always
00:30:42.870 --> 00:30:45.430
have like non like some loss on your
00:30:45.430 --> 00:30:45.890
training set.
00:30:47.970 --> 00:30:50.120
The test error is if you take that same
00:30:50.120 --> 00:30:52.770
model, but now you evaluate it on other
00:30:52.770 --> 00:30:54.516
samples from the distribution, other
00:30:54.516 --> 00:30:56.040
test samples, and you compute an
00:30:56.040 --> 00:30:56.835
expected error.
00:30:56.835 --> 00:30:59.264
The average error over those test
00:30:59.264 --> 00:31:01.172
samples, your test error.
00:31:01.172 --> 00:31:03.330
You always expect your test error to be
00:31:03.330 --> 00:31:04.480
higher than your training error.
00:31:05.130 --> 00:31:06.400
Because you're.
00:31:06.490 --> 00:31:07.000
Time.
00:31:07.860 --> 00:31:10.140
Because your test error was not used to
00:31:10.140 --> 00:31:11.530
optimize your model, but your training
00:31:11.530 --> 00:31:12.000
error was.
00:31:13.140 --> 00:31:15.260
In that gap between the test air and
00:31:15.260 --> 00:31:16.260
the training error is the
00:31:16.260 --> 00:31:17.320
generalization error.
00:31:18.050 --> 00:31:20.560
So that's how that's the error due to
00:31:20.560 --> 00:31:23.680
the challenge of making predictions
00:31:23.680 --> 00:31:25.330
about new samples that were not made in
00:31:25.330 --> 00:31:25.710
training.
00:31:26.340 --> 00:31:27.510
That were not seen in training.
00:31:29.880 --> 00:31:30.260
Question.
00:31:33.240 --> 00:31:35.950
So overfit means that.
00:31:35.950 --> 00:31:37.920
So this isn't the ideal plot for
00:31:37.920 --> 00:31:38.610
overfitting, but.
00:31:39.500 --> 00:31:41.520
Overfitting is that as your model gets
00:31:41.520 --> 00:31:43.600
more complicated, your training error
00:31:43.600 --> 00:31:45.115
will always should always go down.
00:31:45.115 --> 00:31:48.510
You would expect it to go down if you.
00:31:49.070 --> 00:31:52.200
If you, for example were to keep adding
00:31:52.200 --> 00:31:55.040
features to your model, then the same
00:31:55.040 --> 00:31:57.030
model should keep getting better on
00:31:57.030 --> 00:31:58.550
your training set because you've got
00:31:58.550 --> 00:32:00.235
more features with which to fit your
00:32:00.235 --> 00:32:00.810
training data.
00:32:02.050 --> 00:32:04.430
And maybe for a while your test error
00:32:04.430 --> 00:32:06.320
will also go down because you genuinely
00:32:06.320 --> 00:32:07.350
get a better predictor.
00:32:08.190 --> 00:32:10.200
But then at some point, as you continue
00:32:10.200 --> 00:32:12.500
to increase the complexity, the test
00:32:12.500 --> 00:32:13.880
error will start going up.
00:32:13.880 --> 00:32:15.260
Even though the training error keeps
00:32:15.260 --> 00:32:17.540
going down, the test error goes up, and
00:32:17.540 --> 00:32:18.690
that's the point at which you've
00:32:18.690 --> 00:32:19.180
overfit.
00:32:19.920 --> 00:32:21.604
So you can't.
00:32:21.604 --> 00:32:22.165
Really.
00:32:22.165 --> 00:32:24.600
Common, really common conceptual
00:32:24.600 --> 00:32:27.500
mistake that people make is to think
00:32:27.500 --> 00:32:29.670
that once you're training error is 0,
00:32:29.670 --> 00:32:30.890
then you've overfit.
00:32:30.890 --> 00:32:32.060
That's not overfitting.
00:32:32.060 --> 00:32:32.515
You can't.
00:32:32.515 --> 00:32:33.930
You can't look at your training error
00:32:33.930 --> 00:32:35.789
by itself to say that you've overfit.
00:32:36.560 --> 00:32:38.430
Overfitting is when your test error
00:32:38.430 --> 00:32:40.480
starts to go up after increasing the
00:32:40.480 --> 00:32:41.190
complexity.
00:32:43.380 --> 00:32:44.950
So in your homework 2.
00:32:45.850 --> 00:32:47.778
Trees are like a really good way to
00:32:47.778 --> 00:32:49.235
look at overfitting because the
00:32:49.235 --> 00:32:51.280
complexity is like the depth of the
00:32:51.280 --> 00:32:52.983
tree or the number of nodes in the
00:32:52.983 --> 00:32:53.329
tree.
00:32:53.330 --> 00:32:56.530
So in your in your homework two, you're
00:32:56.530 --> 00:32:58.930
going to look at overfitting and how
00:32:58.930 --> 00:33:01.170
the training and test error varies as
00:33:01.170 --> 00:33:02.510
you increase the complexity of your
00:33:02.510 --> 00:33:03.080
classifiers.
00:33:04.230 --> 00:33:04.550
Question.
00:33:09.440 --> 00:33:09.880
Right.
00:33:09.880 --> 00:33:10.820
Yeah, that's a good point.
00:33:10.820 --> 00:33:13.380
So increasing the sample size does not
00:33:13.380 --> 00:33:15.610
Causeway overfitting, but you will
00:33:15.610 --> 00:33:21.280
always get, you should expect to get a
00:33:21.280 --> 00:33:24.070
better fit to the true model, a closer
00:33:24.070 --> 00:33:25.450
fit to the true model as you increase
00:33:25.450 --> 00:33:26.340
the training size.
00:33:26.340 --> 00:33:28.550
The reason that I say I keep on saying
00:33:28.550 --> 00:33:31.860
expect and what that means is that if
00:33:31.860 --> 00:33:34.416
you were to resample this problem, like
00:33:34.416 --> 00:33:36.430
resample your data over and over again.
00:33:36.590 --> 00:33:39.152
Than on average this will happen, but
00:33:39.152 --> 00:33:41.289
in any particular scenario you can get
00:33:41.290 --> 00:33:41.840
unlucky.
00:33:41.840 --> 00:33:44.270
You could add like 5 training examples
00:33:44.270 --> 00:33:46.490
and they're really non representative
00:33:46.490 --> 00:33:48.620
by chance and they cause your model to
00:33:48.620 --> 00:33:49.500
get worse.
00:33:49.500 --> 00:33:51.080
So there's no guarantees.
00:33:51.080 --> 00:33:53.365
But you can say more easily what will
00:33:53.365 --> 00:33:55.980
happen in expectation, which means on
00:33:55.980 --> 00:33:58.420
average under the same kinds of
00:33:58.420 --> 00:33:59.100
situations.
00:34:06.160 --> 00:34:10.527
Alright, so I want to so a lot of a lot
00:34:10.527 --> 00:34:13.120
of people said that these a lot of
00:34:13.120 --> 00:34:14.729
respondents to the survey said that.
00:34:16.090 --> 00:34:17.850
Even when these concepts feel like they
00:34:17.850 --> 00:34:20.910
make sense abstractly or theoretically,
00:34:20.910 --> 00:34:22.540
it's not that easy to understand.
00:34:22.540 --> 00:34:23.749
How do you actually put it into
00:34:23.750 --> 00:34:25.660
practice and turn it into code?
00:34:25.660 --> 00:34:27.750
So I want to work through a particular
00:34:27.750 --> 00:34:29.200
example in some detail.
00:34:30.090 --> 00:34:33.490
And the example I choose is this
00:34:33.490 --> 00:34:35.550
Wisconsin breast cancer data set.
00:34:36.450 --> 00:34:38.290
So this data set was collected in the
00:34:38.290 --> 00:34:39.360
early 90s.
00:34:40.440 --> 00:34:44.650
The motivation is that is that doctors
00:34:44.650 --> 00:34:46.800
wanted to use this tool, called fine
00:34:46.800 --> 00:34:50.410
needle aspirates to diagnose whether a
00:34:50.410 --> 00:34:52.660
tumor is malignant or benign.
00:34:53.900 --> 00:34:54.900
And doctors.
00:34:54.900 --> 00:34:57.040
In some medical papers, doctors
00:34:57.040 --> 00:35:01.360
reported a 94% accuracy in making this
00:35:01.360 --> 00:35:02.540
diagnosis.
00:35:02.540 --> 00:35:06.560
But the authors of this study, the
00:35:06.560 --> 00:35:08.520
first author, is a medical doctor
00:35:08.520 --> 00:35:08.980
himself.
00:35:11.150 --> 00:35:12.490
Have like 2 concerns.
00:35:12.490 --> 00:35:14.210
One is that they want to see if you can
00:35:14.210 --> 00:35:15.327
get a better accuracy.
00:35:15.327 --> 00:35:17.983
They want two or three, maybe they want
00:35:17.983 --> 00:35:19.560
to reduce the amount of expertise
00:35:19.560 --> 00:35:21.160
that's needed in order to make a good
00:35:21.160 --> 00:35:21.925
diagnosis.
00:35:21.925 --> 00:35:24.080
And third, they suspect that these
00:35:24.080 --> 00:35:26.620
reports may be biased because there's a
00:35:26.620 --> 00:35:29.065
they note that there tends to be like a
00:35:29.065 --> 00:35:30.900
bias towards positive results that are.
00:35:30.900 --> 00:35:34.638
I mean, yeah, there tends to be a bias
00:35:34.638 --> 00:35:36.879
towards positive results and reports,
00:35:36.880 --> 00:35:37.130
right?
00:35:37.990 --> 00:35:40.140
People are more likely to report
00:35:40.140 --> 00:35:41.436
something if they think it's good, then
00:35:41.436 --> 00:35:43.240
if they get a disappointing outcome.
00:35:44.810 --> 00:35:47.190
So they want to create computer based
00:35:47.190 --> 00:35:49.250
tests that are less objective and
00:35:49.250 --> 00:35:51.270
provide an effective diagnostic tool.
00:35:52.830 --> 00:35:55.350
So they collected data from 569
00:35:55.350 --> 00:35:58.660
patients and then for developing the
00:35:58.660 --> 00:36:00.584
algorithm and doing their first tests
00:36:00.584 --> 00:36:02.525
and then they collected an additional
00:36:02.525 --> 00:36:03.250
54.
00:36:03.960 --> 00:36:06.570
Data from another 54 patients for their
00:36:06.570 --> 00:36:07.290
final tests.
00:36:08.850 --> 00:36:13.080
And so you can it's like important to
00:36:13.080 --> 00:36:16.090
understand like how painstaking this
00:36:16.090 --> 00:36:18.340
process is of collecting data.
00:36:18.340 --> 00:36:18.740
So.
00:36:19.470 --> 00:36:21.620
These are these are real people who
00:36:21.620 --> 00:36:24.350
have tumors and they take medical
00:36:24.350 --> 00:36:26.660
images of them and then they have some
00:36:26.660 --> 00:36:28.730
interface where somebody can go in and
00:36:28.730 --> 00:36:31.176
outline several of the cells, many of
00:36:31.176 --> 00:36:32.530
the cells that were detected.
00:36:33.930 --> 00:36:35.836
And then they have a.
00:36:35.836 --> 00:36:38.220
Then they do like an automated analysis
00:36:38.220 --> 00:36:40.060
of those outlines to compute different
00:36:40.060 --> 00:36:42.100
features, like how what is the radius
00:36:42.100 --> 00:36:43.853
of the cells and what's the area of the
00:36:43.853 --> 00:36:45.250
cells and what's the compactness.
00:36:46.420 --> 00:36:47.350
And then?
00:36:47.450 --> 00:36:48.110
00:36:48.860 --> 00:36:51.460
As the final features, they look at
00:36:51.460 --> 00:36:53.790
these characteristics of the cells.
00:36:53.790 --> 00:36:54.810
They look at the average
00:36:54.810 --> 00:36:57.162
characteristic, the characteristic of
00:36:57.162 --> 00:36:59.620
the largest cell, the worst cell.
00:37:00.340 --> 00:37:04.030
And the and then the standard deviation
00:37:04.030 --> 00:37:05.340
of these characteristics.
00:37:05.340 --> 00:37:06.730
So they're looking at trying to look at
00:37:06.730 --> 00:37:09.250
like the distribution of these shape
00:37:09.250 --> 00:37:11.680
properties of the cells in order to
00:37:11.680 --> 00:37:13.410
determine if the cancerous cells are
00:37:13.410 --> 00:37:14.390
malignant or benign.
00:37:15.880 --> 00:37:18.820
So it's a pretty involved process to
00:37:18.820 --> 00:37:19.620
collect that data.
00:37:22.080 --> 00:37:22.420
00:38:00.720 --> 00:38:01.480
Right.
00:38:01.480 --> 00:38:04.120
So what you would do?
00:38:04.120 --> 00:38:08.160
And if you go for any kinds of tests,
00:38:08.160 --> 00:38:10.000
you'll probably experience this to some
00:38:10.000 --> 00:38:10.320
extent.
00:38:11.820 --> 00:38:13.870
Like often, somebody will go, a
00:38:13.870 --> 00:38:16.093
technician will go in, they see some
00:38:16.093 --> 00:38:17.710
image, they take different measurements
00:38:17.710 --> 00:38:18.350
on the image.
00:38:19.090 --> 00:38:22.410
And then they can say then they may run
00:38:22.410 --> 00:38:24.765
this like through some data analysis,
00:38:24.765 --> 00:38:27.650
and either either they have rules in
00:38:27.650 --> 00:38:29.640
their head for like what are acceptable
00:38:29.640 --> 00:38:32.715
variations, or they run it through some
00:38:32.715 --> 00:38:36.760
analysis and they'll say, they might
00:38:36.760 --> 00:38:39.110
tell you have no cause for concern, or
00:38:39.110 --> 00:38:41.474
there's like some cause for concern, or
00:38:41.474 --> 00:38:43.350
like there's great cause for concern.
00:38:44.140 --> 00:38:45.510
But if you have an algorithm that it
00:38:45.510 --> 00:38:47.100
might tell you, in this case, for
00:38:47.100 --> 00:38:49.630
example, what's the probability that
00:38:49.630 --> 00:38:51.850
these cells are malignant versus
00:38:51.850 --> 00:38:52.980
benign?
00:38:52.980 --> 00:38:55.595
And then you might say, if there's a
00:38:55.595 --> 00:38:57.730
30% chance that it's malignant, then
00:38:57.730 --> 00:38:59.210
I'm going to recommend a biopsy.
00:38:59.210 --> 00:39:02.160
So you want to have some confidence
00:39:02.160 --> 00:39:03.140
with your prediction.
00:39:04.210 --> 00:39:05.360
So in this.
00:39:06.760 --> 00:39:08.392
In our analysis, we're not going to
00:39:08.392 --> 00:39:11.020
look at the confidences too much for
00:39:11.020 --> 00:39:12.010
simplicity.
00:39:12.010 --> 00:39:15.457
But in the study they also will look,
00:39:15.457 --> 00:39:18.340
they also look at the like specificity,
00:39:18.340 --> 00:39:20.560
like how often can you do you
00:39:20.560 --> 00:39:22.406
misdiagnose one way or the other and
00:39:22.406 --> 00:39:24.155
they can use the confidence as part of
00:39:24.155 --> 00:39:24.860
the recommendation.
00:39:30.410 --> 00:39:35.273
Alright, so I'm going to go into this
00:39:35.273 --> 00:39:37.140
and I think now is a good time to take
00:39:37.140 --> 00:39:38.050
a minute or two.
00:39:38.050 --> 00:39:39.515
You can think about this problem, how
00:39:39.515 --> 00:39:40.250
you would solve it.
00:39:40.250 --> 00:39:42.130
You've got 30 features, continuous
00:39:42.130 --> 00:39:43.510
features, and you're trying to predict
00:39:43.510 --> 00:39:44.450
malignant or benign.
00:39:45.150 --> 00:39:48.480
And also feel free to stretch your it.
00:39:48.480 --> 00:39:51.920
You need to prepare your mind for the
00:39:51.920 --> 00:39:52.410
next half.
00:40:20.140 --> 00:40:20.570
Question.
00:40:36.560 --> 00:40:39.556
Decision trees for example does that
00:40:39.556 --> 00:40:42.250
and neural networks will also do that.
00:40:42.250 --> 00:40:44.940
Or kernelized SVMS and nearest
00:40:44.940 --> 00:40:45.374
neighbor.
00:40:45.374 --> 00:40:47.950
They all they all depend jointly on the
00:40:47.950 --> 00:40:48.560
features.
00:40:51.930 --> 00:40:52.700
How does what?
00:40:56.030 --> 00:40:58.985
I guess because the distance is.
00:40:58.985 --> 00:41:01.517
That's a good point, yeah.
00:41:01.517 --> 00:41:04.160
The K&NI guess, it depends jointly on
00:41:04.160 --> 00:41:05.790
them, but it's independently
00:41:05.790 --> 00:41:07.020
considering those features.
00:41:07.020 --> 00:41:08.180
That's right, yeah.
00:41:20.030 --> 00:41:23.680
But it's nice if it's often hard to
00:41:23.680 --> 00:41:25.810
know what's relevant, and so it's nice.
00:41:25.810 --> 00:41:27.510
The ideal is that you can just collect
00:41:27.510 --> 00:41:28.840
a lot of things that you think might be
00:41:28.840 --> 00:41:30.950
relevant and feed it into the algorithm
00:41:30.950 --> 00:41:34.578
and not have to manually like manually
00:41:34.578 --> 00:41:36.640
like prune it and out.
00:41:42.050 --> 00:41:45.256
Yeah, so one is robust to irrelevant
00:41:45.256 --> 00:41:47.780
features, but if you do L2, it's not so
00:41:47.780 --> 00:41:49.340
robust to irrelevant features.
00:41:49.340 --> 00:41:50.900
So that's like another property of the
00:41:50.900 --> 00:41:52.160
algorithm is whether it has that
00:41:52.160 --> 00:41:52.660
robustness.
00:41:57.120 --> 00:41:59.780
Alright, so let me zoom in a little
00:41:59.780 --> 00:42:00.280
bit.
00:42:03.050 --> 00:42:04.260
I guess over here.
00:42:10.690 --> 00:42:13.660
So we've got this data set.
00:42:13.660 --> 00:42:15.710
Fortunately, in this case, I can load
00:42:15.710 --> 00:42:17.900
the data set from sklearn datasets.
00:42:19.720 --> 00:42:22.300
So here I have the initialization code
00:42:22.300 --> 00:42:22.965
and your homework.
00:42:22.965 --> 00:42:24.790
I provided this code to you as well
00:42:24.790 --> 00:42:26.670
that initially like loads the data and
00:42:26.670 --> 00:42:28.480
splits it up into different datasets.
00:42:29.440 --> 00:42:32.010
But here I've just got my libraries
00:42:32.010 --> 00:42:33.470
that I'm going to use.
00:42:33.470 --> 00:42:37.960
I load the data I this data comes in
00:42:37.960 --> 00:42:39.260
like a particular structure.
00:42:39.260 --> 00:42:40.930
So I take out the features which are
00:42:40.930 --> 00:42:43.740
capital X, the predictions which are Y.
00:42:44.490 --> 00:42:45.940
And it also gives me names of the
00:42:45.940 --> 00:42:49.120
features and names of the predictions
00:42:49.120 --> 00:42:50.690
which are good for visualization.
00:42:51.740 --> 00:42:53.330
So if I run this, it's going to start
00:42:53.330 --> 00:42:55.328
an instance on collabs and then it's
00:42:55.328 --> 00:42:57.366
going to download the data and print
00:42:57.366 --> 00:42:59.900
out the shape and the shape of Y.
00:42:59.900 --> 00:43:02.950
So I often like I print a lot of shapes
00:43:02.950 --> 00:43:05.130
of variables when I'm doing stuff
00:43:05.130 --> 00:43:07.880
because it helps me to make sure I
00:43:07.880 --> 00:43:09.230
understand exactly what I loaded.
00:43:09.230 --> 00:43:11.679
Like if I print out the shape and it's
00:43:11.679 --> 00:43:14.006
if the shape of X is 1 by something
00:43:14.006 --> 00:43:15.660
then I would be like maybe I took the
00:43:15.660 --> 00:43:18.160
wrong like values from this data
00:43:18.160 --> 00:43:18.630
structure.
00:43:19.760 --> 00:43:23.580
Alright, so I've got 569 data points.
00:43:23.580 --> 00:43:26.950
So remember that there were 569 samples
00:43:26.950 --> 00:43:28.790
that were drawn at first that were used
00:43:28.790 --> 00:43:30.350
for their training and algorithm
00:43:30.350 --> 00:43:32.680
development, and then another like 56
00:43:32.680 --> 00:43:34.340
or something that we use for testing.
00:43:34.340 --> 00:43:36.380
The 56 are not released, they're not
00:43:36.380 --> 00:43:37.170
part of this data set.
00:43:38.230 --> 00:43:40.150
And then there's 30 features, there's
00:43:40.150 --> 00:43:41.300
10 characteristics.
00:43:41.970 --> 00:43:44.560
That correspond to the like the worst
00:43:44.560 --> 00:43:46.230
case, the average case and the steering
00:43:46.230 --> 00:43:46.760
deviation.
00:43:47.470 --> 00:43:50.034
And I've got 569 labels, so number of
00:43:50.034 --> 00:43:52.010
labels equals number of data points, so
00:43:52.010 --> 00:43:52.500
that's good.
00:43:54.430 --> 00:43:56.433
Now I can print out.
00:43:56.433 --> 00:43:58.960
I usually will also like print out some
00:43:58.960 --> 00:44:00.940
examples just to make sure that there's
00:44:00.940 --> 00:44:01.585
nothing weird here.
00:44:01.585 --> 00:44:04.125
I don't have any nins or anything like
00:44:04.125 --> 00:44:04.330
that.
00:44:05.190 --> 00:44:06.620
So here are the different feature
00:44:06.620 --> 00:44:08.060
names.
00:44:08.060 --> 00:44:11.080
Here's I chose a few random example
00:44:11.080 --> 00:44:11.760
indices.
00:44:12.430 --> 00:44:14.980
And I can see, I can see some of the
00:44:14.980 --> 00:44:15.740
feature values.
00:44:15.740 --> 00:44:18.530
So there's no NANS or Memphis or
00:44:18.530 --> 00:44:19.570
anything like that in there.
00:44:19.570 --> 00:44:20.400
That's good.
00:44:20.400 --> 00:44:22.320
Also I can notice like.
00:44:23.030 --> 00:44:25.974
Some of some of their values are like
00:44:25.974 --> 00:44:30.416
1.2 E 2 or 11 E 3, so this is like
00:44:30.416 --> 00:44:32.080
1000, while some other ones are really
00:44:32.080 --> 00:44:36.134
small, like 1188 E -, 1.
00:44:36.134 --> 00:44:37.910
So that's something to consider.
00:44:37.910 --> 00:44:39.340
There's a pretty big range of the
00:44:39.340 --> 00:44:40.230
feature values here.
00:44:43.520 --> 00:44:45.600
So then another thing I'll do early is
00:44:45.600 --> 00:44:48.050
say how common is each class, because
00:44:48.050 --> 00:44:50.120
if like 99% of the examples are in one
00:44:50.120 --> 00:44:51.745
class, that's something I need to keep
00:44:51.745 --> 00:44:53.840
in mind versus a 5050 split.
00:44:55.290 --> 00:44:56.650
So in this case.
00:44:56.750 --> 00:44:57.360
00:44:58.700 --> 00:45:02.810
37% of the examples have Class 0 and
00:45:02.810 --> 00:45:04.600
63% have Class 1.
00:45:05.630 --> 00:45:10.190
And if I think I printed the label
00:45:10.190 --> 00:45:12.105
names, yeah, so the label names.
00:45:12.105 --> 00:45:14.750
So 0 means malignant and one means
00:45:14.750 --> 00:45:15.260
benign.
00:45:15.940 --> 00:45:20.190
So in this sample, 37% are malignant
00:45:20.190 --> 00:45:21.940
and 63% are benign.
00:45:24.410 --> 00:45:26.060
Now I'm going to create a training and
00:45:26.060 --> 00:45:27.160
validation set.
00:45:27.160 --> 00:45:29.410
So I define the number of training
00:45:29.410 --> 00:45:31.720
samples 469.
00:45:32.650 --> 00:45:35.845
I use a random seed and that's because
00:45:35.845 --> 00:45:38.360
it might be that the training samples
00:45:38.360 --> 00:45:40.141
are stored in some structured way.
00:45:40.141 --> 00:45:42.125
Maybe they put all the examples with
00:45:42.125 --> 00:45:44.260
zero first, label zero first and then
00:45:44.260 --> 00:45:45.280
label one.
00:45:45.280 --> 00:45:47.629
Or maybe they were structured in some
00:45:47.630 --> 00:45:49.910
other way and I want it to be random,
00:45:49.910 --> 00:45:51.800
so randomness is not something you can
00:45:51.800 --> 00:45:52.720
leave to chance.
00:45:52.720 --> 00:45:56.250
You need to use some permutation to
00:45:56.250 --> 00:45:58.450
make sure that you get a random sample
00:45:58.450 --> 00:45:59.040
of the data.
00:46:00.580 --> 00:46:03.319
So I do a random permutation of the
00:46:03.320 --> 00:46:05.840
same length as the number of indices.
00:46:05.840 --> 00:46:08.280
I set a seed here because I just wanted
00:46:08.280 --> 00:46:10.010
this to be repeatable for the purpose
00:46:10.010 --> 00:46:11.890
of the class, and actually it's a good
00:46:11.890 --> 00:46:14.310
idea to set a seed anyway so that.
00:46:16.450 --> 00:46:18.540
Because takes out one source of
00:46:18.540 --> 00:46:20.210
variance for your debugging.
00:46:21.980 --> 00:46:24.145
So I split it into a training set.
00:46:24.145 --> 00:46:25.770
I took the first untrained.
00:46:26.750 --> 00:46:29.290
It's my X train and Y train and then I
00:46:29.290 --> 00:46:32.420
took all the rest as my X value, Y Val
00:46:32.420 --> 00:46:34.410
and by the 1st examples I mean the
00:46:34.410 --> 00:46:36.060
first ones that in this random
00:46:36.060 --> 00:46:37.130
permutation list.
00:46:38.330 --> 00:46:41.580
Now X train and Y train have.
00:46:42.020 --> 00:46:47.790
I have 469 examples so 469 by 30.
00:46:48.680 --> 00:46:51.575
And X value Y Val which is the second
00:46:51.575 --> 00:46:53.310
one has 100 examples.
00:46:55.420 --> 00:46:58.375
Sometimes the first thing I'll do is
00:46:58.375 --> 00:47:01.360
like a simple classifier just to see is
00:47:01.360 --> 00:47:02.390
this problem trivial.
00:47:02.390 --> 00:47:04.125
If I get like 0 error right away, then
00:47:04.125 --> 00:47:06.780
I can just like stop spend time on it.
00:47:07.630 --> 00:47:10.909
So I made a nearest neighbor
00:47:10.910 --> 00:47:11.620
classifier.
00:47:11.620 --> 00:47:13.390
So I have nearest neighbor.
00:47:13.390 --> 00:47:15.600
X train and Y train are fed in as well
00:47:15.600 --> 00:47:16.340
as X test.
00:47:17.640 --> 00:47:21.470
Pre initialize my predictions, so I do
00:47:21.470 --> 00:47:23.560
initialize it with zeros.
00:47:23.560 --> 00:47:25.990
For each test sample, I take the
00:47:25.990 --> 00:47:27.540
difference from the test sample and all
00:47:27.540 --> 00:47:29.140
the training samples.
00:47:29.140 --> 00:47:30.940
Under the hood, Numpy we'll do like
00:47:30.940 --> 00:47:32.800
broadcasting, which means it will copy
00:47:32.800 --> 00:47:36.085
this as necessary so that the X test
00:47:36.085 --> 00:47:38.669
will be a 1 by 30 and it will copy it
00:47:38.669 --> 00:47:42.560
so that it becomes a 469 by 30.
00:47:43.860 --> 00:47:45.139
Then I take the difference.
00:47:45.140 --> 00:47:46.330
It will be the difference of each
00:47:46.330 --> 00:47:49.270
element of the features and samples.
00:47:49.960 --> 00:47:51.840
Square it will be the square of each
00:47:51.840 --> 00:47:54.660
element and then I sum over axis one
00:47:54.660 --> 00:47:55.920
which is the 2nd axis.
00:47:55.920 --> 00:47:57.210
Zero is the first axis.
00:47:58.110 --> 00:47:59.790
So this will be the sum squared
00:47:59.790 --> 00:48:00.830
distance of the features.
00:48:01.890 --> 00:48:02.770
Euclidean distance.
00:48:02.770 --> 00:48:04.390
You would also take the square root,
00:48:04.390 --> 00:48:05.857
but I don't need to take the square
00:48:05.857 --> 00:48:09.008
root because the minimum of the square
00:48:09.008 --> 00:48:11.104
of a value is the same as the minimum
00:48:11.104 --> 00:48:13.251
of the square of the square of the
00:48:13.251 --> 00:48:13.519
value.
00:48:13.680 --> 00:48:13.890
Right.
00:48:16.060 --> 00:48:19.780
J is the argument distance, so I say J
00:48:19.780 --> 00:48:21.495
equals the argument and this distance.
00:48:21.495 --> 00:48:23.130
So this will give me the index that had
00:48:23.130 --> 00:48:24.010
the minimum distance.
00:48:24.700 --> 00:48:26.420
If I needed more than one, I could use
00:48:26.420 --> 00:48:29.500
argsort and then take like the first K
00:48:29.500 --> 00:48:30.050
indices.
00:48:31.000 --> 00:48:33.386
I assign the test to the training to
00:48:33.386 --> 00:48:34.720
the training sample that had the
00:48:34.720 --> 00:48:36.500
minimum distance and I returned it.
00:48:36.500 --> 00:48:39.240
So nearest neighbor is pretty simple.
00:48:40.800 --> 00:48:43.980
This like if you're a proficient coder,
00:48:43.980 --> 00:48:46.410
it's like a two minutes or whatever to
00:48:46.410 --> 00:48:46.790
decode it.
00:48:48.690 --> 00:48:52.140
Then I'm going to test it, so I then do
00:48:52.140 --> 00:48:54.050
the prediction on the validation set.
00:48:54.050 --> 00:48:55.230
Remember, nearest neighbor has no
00:48:55.230 --> 00:48:56.870
training, so I have no training code
00:48:56.870 --> 00:48:58.105
here, it's just really a prediction
00:48:58.105 --> 00:48:58.430
code.
00:48:59.450 --> 00:49:02.320
And now compute my average accuracy,
00:49:02.320 --> 00:49:05.309
which is why is the number of times the
00:49:05.310 --> 00:49:08.500
mean times that the validation label is
00:49:08.500 --> 00:49:09.760
equal to the prediction label.
00:49:10.710 --> 00:49:12.230
And then the error is 1 minus the
00:49:12.230 --> 00:49:13.490
accuracy, right?
00:49:13.490 --> 00:49:14.040
So let's run it.
00:49:16.480 --> 00:49:21.550
All right, so I got an error of 8% now.
00:49:23.090 --> 00:49:24.060
I could quit here.
00:49:24.060 --> 00:49:26.840
I could be like, OK, I'm done 8%, but I
00:49:26.840 --> 00:49:28.150
shouldn't really be satisfied with
00:49:28.150 --> 00:49:29.080
this, right?
00:49:29.080 --> 00:49:32.400
So the remember that in the study they
00:49:32.400 --> 00:49:34.105
said that doctors were reporting that
00:49:34.105 --> 00:49:37.380
they can get like 6% error, they had
00:49:37.380 --> 00:49:38.810
94% accuracy.
00:49:39.530 --> 00:49:41.906
And since I'm a machine learning
00:49:41.906 --> 00:49:43.940
machine learning engineer, I'm armed
00:49:43.940 --> 00:49:44.800
with data.
00:49:44.800 --> 00:49:47.250
I should be able to outperform a
00:49:47.250 --> 00:49:49.190
medical Doctor Who has years of
00:49:49.190 --> 00:49:51.960
experience on the same problem.
00:49:54.860 --> 00:49:56.800
Right, so all of his wits and
00:49:56.800 --> 00:49:58.420
experience is just bringing a knife to
00:49:58.420 --> 00:49:59.300
a gunfight.
00:50:01.760 --> 00:50:02.410
I'm just kidding.
00:50:03.810 --> 00:50:05.670
But seriously, like, I can probably do
00:50:05.670 --> 00:50:06.130
better, right?
00:50:06.130 --> 00:50:07.190
It's just my first attempt.
00:50:07.900 --> 00:50:09.530
So let's look at the data a little bit
00:50:09.530 --> 00:50:11.440
better, a little more in depth.
00:50:12.340 --> 00:50:13.610
So remember that one thing we noticed
00:50:13.610 --> 00:50:15.145
is that it looked like some feature
00:50:15.145 --> 00:50:16.895
values were a lot larger than other
00:50:16.895 --> 00:50:18.540
values, and nearest neighbor is not
00:50:18.540 --> 00:50:19.716
very robust to that.
00:50:19.716 --> 00:50:22.830
It might be like emphasizing the large
00:50:22.830 --> 00:50:24.620
values much more, which might not be
00:50:24.620 --> 00:50:25.840
the most important features.
00:50:26.490 --> 00:50:28.390
So here I have a print statement.
00:50:28.390 --> 00:50:30.210
The only thing fancy is that I use some
00:50:30.210 --> 00:50:32.900
spacing thing to make it like evenly
00:50:32.900 --> 00:50:33.420
spaced.
00:50:34.040 --> 00:50:35.828
And I'm printing the means of the
00:50:35.828 --> 00:50:37.330
features, the standard deviations of
00:50:37.330 --> 00:50:39.710
the features, the means of the features
00:50:39.710 --> 00:50:42.413
where y = 1 zero, and the means of the
00:50:42.413 --> 00:50:43.599
features were y = 1.
00:50:44.340 --> 00:50:46.250
So that can kind of tell me a couple
00:50:46.250 --> 00:50:46.580
things.
00:50:46.580 --> 00:50:48.100
One is like what is the scale that
00:50:48.100 --> 00:50:49.530
features by looking at the steering
00:50:49.530 --> 00:50:50.310
deviation and the mean.
00:50:51.170 --> 00:50:54.050
Also, are the features like predictive
00:50:54.050 --> 00:50:54.338
or not?
00:50:54.338 --> 00:50:56.315
If I have a good spread of the means of
00:50:56.315 --> 00:50:59.095
the two features, I mean of the of y =
00:50:59.095 --> 00:51:01.749
0 and y = 1, then it's predictive.
00:51:01.750 --> 00:51:03.600
But if I have a small spread compared
00:51:03.600 --> 00:51:05.530
to the steering deviation then it's not
00:51:05.530 --> 00:51:06.240
very predictive.
00:51:07.350 --> 00:51:10.150
Right, so for example, this feature
00:51:10.150 --> 00:51:11.824
here means smoothness.
00:51:11.824 --> 00:51:15.584
Mean is 1, standard deviation is 01,
00:51:15.584 --> 00:51:19.947
the mean of zero is 1, the mean of one
00:51:19.947 --> 00:51:20.690
is 09.
00:51:20.690 --> 00:51:22.770
And you know with three digits there
00:51:22.770 --> 00:51:24.305
might be look even closer.
00:51:24.305 --> 00:51:26.092
So obviously smoothness means
00:51:26.092 --> 00:51:28.430
smoothness is not a very good feature,
00:51:28.430 --> 00:51:31.340
it's not very predictive of the label.
00:51:32.120 --> 00:51:35.050
Where if I look at something like.
00:51:35.140 --> 00:51:35.930
00:51:37.780 --> 00:51:40.240
If I look at something like this, just
00:51:40.240 --> 00:51:42.125
take the first one, the difference of
00:51:42.125 --> 00:51:43.730
the means is more than one steering
00:51:43.730 --> 00:51:47.620
deviation of the feature, and so mean
00:51:47.620 --> 00:51:49.420
radius is like fairly predictive.
00:51:51.210 --> 00:51:53.395
But the thing my take home from this is
00:51:53.395 --> 00:51:56.480
that some features have means and
00:51:56.480 --> 00:51:58.950
standard deviations that are sub one
00:51:58.950 --> 00:51:59.730
less than one.
00:52:00.400 --> 00:52:03.340
And others are in the hundreds, so not
00:52:03.340 --> 00:52:04.540
that's not good.
00:52:04.540 --> 00:52:05.700
So I want to do some kind of
00:52:05.700 --> 00:52:06.590
normalization.
00:52:09.520 --> 00:52:11.857
So I'm going to normalize by the mean
00:52:11.857 --> 00:52:13.820
and steering deviation, which means
00:52:13.820 --> 00:52:16.537
that I subtract the mean and divide by
00:52:16.537 --> 00:52:17.880
the standard deviation.
00:52:17.880 --> 00:52:20.040
Importantly, you want to compute the
00:52:20.040 --> 00:52:22.138
mean and the standard deviation once on
00:52:22.138 --> 00:52:23.880
the training set and then apply the
00:52:23.880 --> 00:52:25.531
same normalization to the training and
00:52:25.531 --> 00:52:26.566
the validation set.
00:52:26.566 --> 00:52:28.580
So you can't provide different
00:52:28.580 --> 00:52:31.620
normalizations to different sets, or
00:52:31.620 --> 00:52:33.080
else you're going to your features will
00:52:33.080 --> 00:52:35.030
not be comparable and you'll it's a
00:52:35.030 --> 00:52:35.640
bug.
00:52:35.640 --> 00:52:37.360
It's so it won't work.
00:52:38.650 --> 00:52:40.240
OK, so I compute the mean compute
00:52:40.240 --> 00:52:41.720
steering, aviation take the difference,
00:52:41.720 --> 00:52:43.160
divide by zero and aviation do the same
00:52:43.160 --> 00:52:44.220
thing on my valve set.
00:52:44.990 --> 00:52:46.430
And there's nothing to print here, but
00:52:46.430 --> 00:52:47.430
I need to run it.
00:52:47.430 --> 00:52:48.000
Whoops.
00:52:51.250 --> 00:52:52.380
All right, so now I'm going to repeat
00:52:52.380 --> 00:52:53.150
my nearest neighbor.
00:52:53.920 --> 00:52:54.866
OK, 4%.
00:52:54.866 --> 00:52:57.336
So there was a lot better before I got
00:52:57.336 --> 00:53:01.206
12%, I think 8%, yeah, so before I got
00:53:01.206 --> 00:53:01.500
8%.
00:53:02.130 --> 00:53:03.200
Now it's 4%.
00:53:04.050 --> 00:53:04.720
So that's good.
00:53:05.380 --> 00:53:07.040
But I still don't know if like nearest
00:53:07.040 --> 00:53:07.850
neighbor is the best.
00:53:07.850 --> 00:53:09.240
So I shouldn't just try like 1
00:53:09.240 --> 00:53:11.140
algorithm and then assume that's the
00:53:11.140 --> 00:53:11.910
best I should get.
00:53:11.910 --> 00:53:14.620
I should try other algorithms and try
00:53:14.620 --> 00:53:16.280
to see if I can improve things further.
00:53:17.510 --> 00:53:18.110
Question.
00:53:24.670 --> 00:53:25.550
So the yes.
00:53:25.550 --> 00:53:26.940
So the question is why did the error
00:53:26.940 --> 00:53:28.170
rate get better?
00:53:28.170 --> 00:53:30.950
And I think it's because under the
00:53:30.950 --> 00:53:33.920
original features, these features like
00:53:33.920 --> 00:53:38.000
mean area that have a huge range are
00:53:38.000 --> 00:53:40.690
going to dominate the distances.
00:53:40.690 --> 00:53:42.420
All of these features concavity,
00:53:42.420 --> 00:53:45.470
compactness, concave point, symmetry at
00:53:45.470 --> 00:53:48.730
mostly we'll add a distance of .1 or
00:53:48.730 --> 00:53:51.010
something like that where this mean
00:53:51.010 --> 00:53:53.887
area is going to tend to add distances
00:53:53.887 --> 00:53:54.430
of.
00:53:54.490 --> 00:53:54.960
Hundreds.
00:53:55.580 --> 00:53:58.620
And so if I don't normalize it, that
00:53:58.620 --> 00:54:00.100
means that essentially I'm seeing the
00:54:00.100 --> 00:54:01.728
bigger the feature values, the more
00:54:01.728 --> 00:54:02.990
important they are, or the more
00:54:02.990 --> 00:54:04.307
variants and the feature values, the
00:54:04.307 --> 00:54:05.049
more important they are.
00:54:05.670 --> 00:54:07.340
And that's not based on any like
00:54:07.340 --> 00:54:08.480
knowledge of the problem.
00:54:08.480 --> 00:54:09.970
That was just because that's how the
00:54:09.970 --> 00:54:10.720
data turned out.
00:54:10.720 --> 00:54:12.560
And so I don't really trust that kind
00:54:12.560 --> 00:54:14.210
of decision.
00:54:16.270 --> 00:54:16.650
Go ahead.
00:54:18.070 --> 00:54:18.350
OK.
00:54:19.290 --> 00:54:20.240
You had a question?
00:54:29.700 --> 00:54:32.490
So I compute the mean and this is
00:54:32.490 --> 00:54:34.615
computing the mean over the first axis.
00:54:34.615 --> 00:54:36.640
So it means that for every feature
00:54:36.640 --> 00:54:38.700
value I compute the mean over all the
00:54:38.700 --> 00:54:39.320
examples.
00:54:40.110 --> 00:54:42.680
Of the training features XTR.
00:54:43.450 --> 00:54:45.560
So I computed the mean, the expectation
00:54:45.560 --> 00:54:49.370
or the arithmetic average of each
00:54:49.370 --> 00:54:50.010
feature.
00:54:50.990 --> 00:54:53.330
Over all the training samples, and then
00:54:53.330 --> 00:54:56.500
I compute this stern deviation of each
00:54:56.500 --> 00:54:58.140
feature over all the examples.
00:54:58.140 --> 00:54:58.960
So that's the.
00:55:00.570 --> 00:55:00.940
Right.
00:55:03.150 --> 00:55:07.200
So remember that X train has this shape
00:55:07.200 --> 00:55:11.920
469 by 30, so if I go down the first
00:55:11.920 --> 00:55:14.480
axis then I'm changing the example.
00:55:14.480 --> 00:55:17.330
So 0123 et cetera are different
00:55:17.330 --> 00:55:18.100
examples.
00:55:18.100 --> 00:55:20.363
And if I go down the second axis then
00:55:20.363 --> 00:55:22.470
I'm going into different feature
00:55:22.470 --> 00:55:22.960
columns.
00:55:23.680 --> 00:55:25.760
And so I want to take the mean over the
00:55:25.760 --> 00:55:27.524
examples for each feature.
00:55:27.524 --> 00:55:30.113
And so I say access equals zero for the
00:55:30.113 --> 00:55:31.870
mean to take the mean over samples.
00:55:31.870 --> 00:55:34.774
Otherwise I'll end up with a 1 by 30
00:55:34.774 --> 00:55:38.480
where I mean with a 469 by 1 where I've
00:55:38.480 --> 00:55:39.850
taken the average feature for each
00:55:39.850 --> 00:55:40.380
example.
00:55:46.980 --> 00:55:49.390
So if I say X is equals zero, it means
00:55:49.390 --> 00:55:51.000
it will take the mean over all the
00:55:51.000 --> 00:55:52.400
remaining dimensions.
00:55:52.750 --> 00:55:53.320
And.
00:55:54.040 --> 00:55:55.590
Averaging over the first dimension.
00:56:02.380 --> 00:56:04.870
So then this will be a 30 dimensional
00:56:04.870 --> 00:56:06.080
vector X MU.
00:56:07.050 --> 00:56:11.230
It will be the mean of each feature
00:56:11.230 --> 00:56:12.060
over the samples.
00:56:12.930 --> 00:56:14.540
And this is also a 30 dimensional
00:56:14.540 --> 00:56:15.880
vector standard deviation.
00:56:17.170 --> 00:56:19.300
And then I'm subtracting off the mean
00:56:19.300 --> 00:56:21.185
and dividing by the standard deviation.
00:56:21.185 --> 00:56:24.150
And Numpy is nice that even though X
00:56:24.150 --> 00:56:28.355
train is 469 by 30 and X mu is 30, is
00:56:28.355 --> 00:56:28.840
30.
00:56:29.030 --> 00:56:32.370
Numpy is smart, and it says you're
00:56:32.370 --> 00:56:35.390
doing a 469 by 30 -, A thirty.
00:56:35.390 --> 00:56:39.060
So I need to copy that 3469 times to
00:56:39.060 --> 00:56:39.810
take the difference.
00:56:41.550 --> 00:56:42.790
And same for the divide.
00:56:42.790 --> 00:56:44.990
This is an element wise divide so it's
00:56:44.990 --> 00:56:45.800
important to know.
00:56:46.500 --> 00:56:48.340
There you can have like a matrix
00:56:48.340 --> 00:56:50.710
multiplication or matrix inverse or you
00:56:50.710 --> 00:56:53.006
can have an element wise multiplication
00:56:53.006 --> 00:56:53.759
or inverse.
00:56:54.570 --> 00:56:57.070
Usually like the simple operators are
00:56:57.070 --> 00:56:58.320
element wise in Python.
00:56:58.970 --> 00:57:01.485
So this means that for every element of
00:57:01.485 --> 00:57:04.796
this matrix, I'm going to divide by the
00:57:04.796 --> 00:57:06.940
standard deviation the corresponding
00:57:06.940 --> 00:57:07.680
standard deviation.
00:57:09.390 --> 00:57:10.690
And then I do the same thing for the
00:57:10.690 --> 00:57:11.640
validation set.
00:57:11.640 --> 00:57:12.960
And what was your question?
00:57:22.780 --> 00:57:23.490
Yeah.
00:57:32.420 --> 00:57:37.550
So L1 used L1 regularization for linear
00:57:37.550 --> 00:57:40.183
logistic regression and that will that
00:57:40.183 --> 00:57:43.110
will like put that will like select
00:57:43.110 --> 00:57:44.030
features for.
00:57:44.030 --> 00:57:46.110
You could also use L1 nearest neighbor
00:57:46.110 --> 00:57:47.720
distance which would be less sensitive
00:57:47.720 --> 00:57:48.110
to this.
00:57:49.700 --> 00:57:52.150
But with this range of like .1 versus
00:57:52.150 --> 00:57:54.590
like 500, it will still be that the
00:57:54.590 --> 00:57:55.820
larger features will dominate.
00:57:57.180 --> 00:57:57.430
Yep.
00:57:59.850 --> 00:58:03.560
All right, so after I normalized, now
00:58:03.560 --> 00:58:06.550
note that I'm passing in X train N,
00:58:06.550 --> 00:58:08.670
which is for stands for norm for me.
00:58:09.450 --> 00:58:10.380
In X Val north.
00:58:10.380 --> 00:58:12.240
Now I get lower error.
00:58:12.830 --> 00:58:14.220
Alright, so now let's try a different
00:58:14.220 --> 00:58:14.885
classifier.
00:58:14.885 --> 00:58:17.340
Let's do Naive Bayes, and I'm going to
00:58:17.340 --> 00:58:21.055
assume that each feature value given
00:58:21.055 --> 00:58:23.399
the class is a Gaussian.
00:58:23.399 --> 00:58:27.480
So given that y = 0, Y equals one.
00:58:27.480 --> 00:58:30.232
Then my probability of the feature is a
00:58:30.232 --> 00:58:31.770
Gaussian with some mean and some
00:58:31.770 --> 00:58:32.680
standard deviation.
00:58:33.410 --> 00:58:35.640
Now for nibs I need a training and
00:58:35.640 --> 00:58:36.610
prediction function.
00:58:37.590 --> 00:58:40.560
So I'm going to pass in my training
00:58:40.560 --> 00:58:41.430
data X&Y.
00:58:42.300 --> 00:58:44.760
App says some like I'm going to use
00:58:44.760 --> 00:58:46.864
that as like a prior to add it to the
00:58:46.864 --> 00:58:48.390
variance so that even if my feature
00:58:48.390 --> 00:58:50.340
value has no variance in training, I'm
00:58:50.340 --> 00:58:52.175
going to have some minimal variance so
00:58:52.175 --> 00:58:54.450
that I don't have like a divide by zero
00:58:54.450 --> 00:58:56.610
essentially where I'm not like over
00:58:56.610 --> 00:59:00.600
relying on the variance that I observe.
00:59:02.080 --> 00:59:03.960
All right, so initialize my MU and my
00:59:03.960 --> 00:59:06.988
Sigma to be the number of features by
00:59:06.988 --> 00:59:08.880
two, and the two is because there's two
00:59:08.880 --> 00:59:10.360
classes, so I'm going to estimate this
00:59:10.360 --> 00:59:10.960
for each class.
00:59:12.250 --> 00:59:14.988
I compute my probability of the label
00:59:14.988 --> 00:59:17.870
to be just the mean of y = 0.
00:59:17.870 --> 00:59:19.180
So this is a probability that the label
00:59:19.180 --> 00:59:20.000
is equal to 0.
00:59:21.530 --> 00:59:23.820
And then for each feature so range,
00:59:23.820 --> 00:59:25.650
you'll be 0 to the number of features.
00:59:26.510 --> 00:59:30.100
I compute the mean over the cases where
00:59:30.100 --> 00:59:31.330
the label equals 0.
00:59:32.660 --> 00:59:34.770
And the mean over the case where the
00:59:34.770 --> 00:59:36.450
labels equals one.
00:59:36.450 --> 00:59:37.990
And I could do this as like a
00:59:37.990 --> 00:59:40.260
vectorized operation like over an axis,
00:59:40.260 --> 00:59:41.970
but for clarity I did it this way.
00:59:42.700 --> 00:59:43.350
With the four loop.
00:59:45.040 --> 00:59:47.990
Compute their stern deviation where y =
00:59:47.990 --> 00:59:50.827
0 and the stereo deviation where y = 1
00:59:50.827 --> 00:59:52.520
and again like this epsilon will be
00:59:52.520 --> 00:59:55.600
some small number that will just like
00:59:55.600 --> 00:59:57.260
make sure that my variance isn't zero.
00:59:57.260 --> 00:59:59.810
Or like says that like I think there
00:59:59.810 --> 01:00:01.030
might be a little bit more variance
01:00:01.030 --> 01:00:01.740
than I observe.
01:00:03.080 --> 01:00:03.600
And.
01:00:04.420 --> 01:00:05.090
That's it.
01:00:05.090 --> 01:00:07.570
So then I'll return my mean steering
01:00:07.570 --> 01:00:09.150
deviation and the probability of the
01:00:09.150 --> 01:00:10.010
label question.
01:00:12.500 --> 01:00:12.760
Sorry.
01:00:21.950 --> 01:00:24.952
Because X shape one, so X shape zero is
01:00:24.952 --> 01:00:26.505
the number of samples and X shape one
01:00:26.505 --> 01:00:27.840
is the number of features.
01:00:27.840 --> 01:00:30.810
And there's a mean for every mean
01:00:30.810 --> 01:00:33.273
estimate for every feature, not for
01:00:33.273 --> 01:00:34.050
every sample.
01:00:35.780 --> 01:00:37.840
So this will be a number of features by
01:00:37.840 --> 01:00:38.230
two.
01:00:43.510 --> 01:00:44.720
Alright, and then I'm going to do
01:00:44.720 --> 01:00:45.380
prediction.
01:00:45.380 --> 01:00:48.200
So now I'll write my prediction code.
01:00:48.200 --> 01:00:50.080
I now need to pass in the thing that I
01:00:50.080 --> 01:00:50.930
want to predict for.
01:00:51.620 --> 01:00:53.720
That means in the steering deviations
01:00:53.720 --> 01:00:55.840
and the P0 that I estimated from my
01:00:55.840 --> 01:00:56.670
training function.
01:00:57.640 --> 01:01:00.450
And I'm going to compute the log
01:01:00.450 --> 01:01:04.460
probability of X given of X&Y, not the
01:01:04.460 --> 01:01:05.390
probability of X&Y.
01:01:06.130 --> 01:01:07.889
And the reason for that is that if I
01:01:07.890 --> 01:01:09.960
multiply a lot of small probabilities
01:01:09.960 --> 01:01:11.706
together then I get a really small
01:01:11.706 --> 01:01:11.972
number.
01:01:11.972 --> 01:01:13.955
And if I have a lot of features like
01:01:13.955 --> 01:01:16.418
you do for MNIST for example, then that
01:01:16.418 --> 01:01:18.470
small number will eventually become
01:01:18.470 --> 01:01:21.820
zero and like in terms of floating
01:01:21.820 --> 01:01:23.889
point operations or it will become like
01:01:23.890 --> 01:01:26.470
unwieldly small.
01:01:26.470 --> 01:01:28.160
So you want to compute the log
01:01:28.160 --> 01:01:29.460
probability, not the probability.
01:01:30.460 --> 01:01:33.100
And minimizing the OR maximizing the
01:01:33.100 --> 01:01:34.602
log probability is the same as
01:01:34.602 --> 01:01:35.660
maximizing the probability.
01:01:36.860 --> 01:01:38.560
So for each feature.
01:01:39.350 --> 01:01:43.388
I add the log probability of the
01:01:43.388 --> 01:01:46.726
feature given y = 0 or the feature
01:01:46.726 --> 01:01:47.739
given y = 1.
01:01:48.960 --> 01:01:53.265
And this is this is the log of the
01:01:53.265 --> 01:01:54.000
Gaussian function.
01:01:54.000 --> 01:01:56.340
Just ignoring the constant multiplier
01:01:56.340 --> 01:01:58.540
in the Gaussian function because that
01:01:58.540 --> 01:02:01.300
won't be any different whether y = 0
01:02:01.300 --> 01:02:03.059
one there one over square root, square
01:02:03.059 --> 01:02:04.310
root 2π is Sigma.
01:02:06.200 --> 01:02:12.750
So this minus mean minus X ^2 divided
01:02:12.750 --> 01:02:14.140
by Sigma squared.
01:02:14.140 --> 01:02:15.930
That's like in the exponent of the
01:02:15.930 --> 01:02:16.490
Gaussian.
01:02:16.490 --> 01:02:18.530
So when I take the log of it, I've just
01:02:18.530 --> 01:02:19.860
got that exponent there.
01:02:20.820 --> 01:02:25.040
So I'm adding that to my score of log
01:02:25.040 --> 01:02:29.630
PX y = 0 and log pxy equals one.
01:02:32.780 --> 01:02:35.721
Then I'm adding my prior so to my 0
01:02:35.721 --> 01:02:38.204
score I add the log probability of y =
01:02:38.204 --> 01:02:38.479
0.
01:02:38.480 --> 01:02:41.440
Into my one score, I add the log
01:02:41.440 --> 01:02:44.230
probability of y = 1, which is just one
01:02:44.230 --> 01:02:45.729
minus the probability of y = 0.
01:02:46.780 --> 01:02:48.540
And then I take the argmax to get my
01:02:48.540 --> 01:02:50.899
prediction and I'm taking the argmax
01:02:50.900 --> 01:02:53.910
over axis one because that was my label
01:02:53.910 --> 01:02:54.380
axis.
01:02:55.170 --> 01:02:55.720
So.
01:02:56.860 --> 01:02:58.875
So here the first axis is the number of
01:02:58.875 --> 01:03:00.915
test samples, the second axis is the
01:03:00.915 --> 01:03:01.860
number of labels.
01:03:01.860 --> 01:03:04.470
I take the argmax over the labels to
01:03:04.470 --> 01:03:07.820
get my maximum my most likely
01:03:07.820 --> 01:03:09.510
prediction for every test sample.
01:03:13.750 --> 01:03:15.930
And then finally the code to call this
01:03:15.930 --> 01:03:18.334
so I call Gaussian train NI Bayes
01:03:18.334 --> 01:03:21.650
Gaussian train and I use this as my as
01:03:21.650 --> 01:03:23.800
like my prior on the variance my
01:03:23.800 --> 01:03:24.290
epsilon.
01:03:25.400 --> 01:03:29.310
And then I'd call predict and I pass in
01:03:29.310 --> 01:03:30.240
the validation data.
01:03:31.200 --> 01:03:32.510
And then I measure my error.
01:03:33.400 --> 01:03:35.130
And I'm going to do this.
01:03:35.130 --> 01:03:36.970
So here's a question.
01:03:36.970 --> 01:03:39.338
Do you think that here I'm doing it on
01:03:39.338 --> 01:03:41.219
the non normalized features and here
01:03:41.219 --> 01:03:43.182
I'm doing it on the normalized
01:03:43.182 --> 01:03:43.509
features?
01:03:44.380 --> 01:03:47.160
Do you think that those results will be
01:03:47.160 --> 01:03:48.800
different or the same?
01:03:48.800 --> 01:03:50.510
So how many people think that these
01:03:50.510 --> 01:03:52.260
will be the same if I?
01:03:52.960 --> 01:03:56.930
Do not have bays on rescaled and mean
01:03:56.930 --> 01:04:00.130
normalized features versus normalized.
01:04:01.370 --> 01:04:02.790
So how many people think it will be the
01:04:02.790 --> 01:04:03.640
same result?
01:04:05.470 --> 01:04:07.060
OK, how many people think it will be a
01:04:07.060 --> 01:04:07.610
different result?
01:04:10.570 --> 01:04:12.250
About 5050.
01:04:12.250 --> 01:04:13.510
Alright, so let's see.
01:04:13.510 --> 01:04:14.820
Let's see how it turns out.
01:04:18.860 --> 01:04:20.980
So it's exactly the same, and it's
01:04:20.980 --> 01:04:22.855
actually guaranteed to be exactly the
01:04:22.855 --> 01:04:25.350
same in this case because.
01:04:27.190 --> 01:04:28.790
Because if I scale or shift the
01:04:28.790 --> 01:04:30.910
features, all it's going to do is
01:04:30.910 --> 01:04:32.320
change my mean invariance.
01:04:32.960 --> 01:04:34.420
But it will change it the same way for
01:04:34.420 --> 01:04:36.500
each class, so the probability of the
01:04:36.500 --> 01:04:38.450
features given the data given the label
01:04:38.450 --> 01:04:40.540
doesn't change at all when I shift them
01:04:40.540 --> 01:04:42.050
or scale them according to a Gaussian
01:04:42.050 --> 01:04:42.990
distribution.
01:04:42.990 --> 01:04:45.080
So that's why the feature normalization
01:04:45.080 --> 01:04:46.790
isn't really necessary here for Naive
01:04:46.790 --> 01:04:47.060
Bayes.
01:04:48.890 --> 01:04:50.605
But it wasn't didn't do great.
01:04:50.605 --> 01:04:51.790
It doesn't usually.
01:04:51.790 --> 01:04:52.870
So not a big surprise.
01:04:54.240 --> 01:04:56.697
So then finally, let's do.
01:04:56.697 --> 01:04:58.500
Let's put in a logistic there.
01:04:58.500 --> 01:05:00.100
Let's do linear and logistic
01:05:00.100 --> 01:05:03.060
regression, and I'm going to use the
01:05:03.060 --> 01:05:03.770
model here.
01:05:04.510 --> 01:05:06.700
So C = 1 is the default that's Lambda
01:05:06.700 --> 01:05:07.650
equals one.
01:05:07.650 --> 01:05:09.410
I'll give it plenty of iterations, just
01:05:09.410 --> 01:05:10.750
make sure it can converge.
01:05:10.750 --> 01:05:12.350
I fit it on the training data.
01:05:13.230 --> 01:05:15.310
Test it on the validation data.
01:05:15.310 --> 01:05:17.270
And here I'm going to compare for if I
01:05:17.270 --> 01:05:19.230
don't normalize versus I normalize.
01:05:23.690 --> 01:05:27.037
And so in this case I got 3% error when
01:05:27.037 --> 01:05:29.907
I didn't normalize and I got 0% error
01:05:29.907 --> 01:05:31.350
when I normalized.
01:05:33.670 --> 01:05:34.990
So the normalization.
01:05:34.990 --> 01:05:36.470
The reason it makes a difference in
01:05:36.470 --> 01:05:39.070
this linear model is that I have some
01:05:39.070 --> 01:05:40.100
regularization weight.
01:05:40.770 --> 01:05:43.420
So if I set this to something really
01:05:43.420 --> 01:05:46.780
big, SK learn is a little awkward and
01:05:46.780 --> 01:05:48.620
that C is the inverse of Lambda.
01:05:48.620 --> 01:05:50.970
So the higher this value is, the less
01:05:50.970 --> 01:05:51.970
the regularization.
01:05:58.010 --> 01:06:00.710
I thought they would do something, but
01:06:00.710 --> 01:06:01.240
it didn't.
01:06:03.440 --> 01:06:05.290
That's not going to make a difference.
01:06:06.730 --> 01:06:07.790
That's interesting actually.
01:06:07.790 --> 01:06:08.970
I don't know why.
01:06:09.730 --> 01:06:11.180
Maybe I maybe I got.
01:06:11.180 --> 01:06:13.510
Let's see, let's make it really small
01:06:13.510 --> 01:06:13.980
instead.
01:06:24.460 --> 01:06:24.920
What's what?
01:06:29.290 --> 01:06:32.130
So that definitely changed things, but
01:06:32.130 --> 01:06:33.620
it made the normalization worse.
01:06:33.620 --> 01:06:34.500
That's interesting.
01:06:34.500 --> 01:06:36.420
OK, I cannot explain that off the dot
01:06:36.420 --> 01:06:37.200
my head.
01:06:38.070 --> 01:06:41.200
But another thing is that if I do 0.
01:06:42.740 --> 01:06:44.095
Wait, actually zero.
01:06:44.095 --> 01:06:46.425
I don't remember again if which way?
01:06:46.425 --> 01:06:47.710
I have to, yeah.
01:06:48.470 --> 01:06:48.990
So.
01:06:50.650 --> 01:06:52.340
You need like you need some
01:06:52.340 --> 01:06:53.280
regularization.
01:06:54.220 --> 01:06:55.780
Or else you get errors like that.
01:06:58.220 --> 01:07:01.460
They're not regularizing is not info.
01:07:02.560 --> 01:07:05.650
Not regularizing is usually not an
01:07:05.650 --> 01:07:05.980
option.
01:07:05.980 --> 01:07:07.070
OK, never mind, all right.
01:07:08.140 --> 01:07:10.723
Yeah, you guys can play with it if you
01:07:10.723 --> 01:07:10.859
want.
01:07:10.860 --> 01:07:11.323
I'm going to.
01:07:11.323 --> 01:07:12.910
I just, I don't want to get stuck there
01:07:12.910 --> 01:07:15.340
as getting too much into the weeds.
01:07:16.530 --> 01:07:20.235
The normalization helped in the case of
01:07:20.235 --> 01:07:22.370
the default regularization.
01:07:24.010 --> 01:07:27.120
I can also plot a.
01:07:27.790 --> 01:07:29.590
I can also do like other ways of
01:07:29.590 --> 01:07:31.360
looking at the data.
01:07:31.360 --> 01:07:32.550
Let's look at.
01:07:32.550 --> 01:07:34.390
I'm going to change this since it was
01:07:34.390 --> 01:07:35.520
kind of boring.
01:07:37.500 --> 01:07:38.410
Let me just.
01:07:38.500 --> 01:07:39.190
01:07:41.150 --> 01:07:41.510
Whoops.
01:07:42.630 --> 01:07:44.430
I don't it's not very interesting to
01:07:44.430 --> 01:07:46.340
look at an Roc curve if you get perfect
01:07:46.340 --> 01:07:46.910
prediction.
01:07:48.670 --> 01:07:50.290
So let me just change this a little
01:07:50.290 --> 01:07:50.640
bit.
01:07:52.040 --> 01:07:54.870
So I'm going to look at the one where I
01:07:54.870 --> 01:07:56.380
did not perfect prediction.
01:07:57.840 --> 01:07:58.650
01:08:00.300 --> 01:08:00.830
Mexican.
01:08:03.700 --> 01:08:07.390
Right, so this arc curve shows me given
01:08:07.390 --> 01:08:09.320
if I choose different thresholds on my
01:08:09.320 --> 01:08:10.000
confidence.
01:08:10.870 --> 01:08:13.535
By default, you choose a confidence at
01:08:13.535 --> 01:08:14.050
5:00.
01:08:14.050 --> 01:08:15.810
If probability is greater than five,
01:08:15.810 --> 01:08:17.810
then you assign it to the class that
01:08:17.810 --> 01:08:19.069
had that greater probability.
01:08:19.700 --> 01:08:21.440
But you can say for example if the
01:08:21.440 --> 01:08:23.820
probability is greater than .3 then I'm
01:08:23.820 --> 01:08:27.030
going to say it's like malignant and
01:08:27.030 --> 01:08:28.150
otherwise it's benign.
01:08:28.150 --> 01:08:29.740
So you can choose different thresholds.
01:08:30.450 --> 01:08:31.990
Especially if there's a different
01:08:31.990 --> 01:08:33.440
consequence to getting either one
01:08:33.440 --> 01:08:36.100
wrong, like which there is for
01:08:36.100 --> 01:08:37.260
malignant versus benign.
01:08:38.080 --> 01:08:40.530
So you can look at this arc curve which
01:08:40.530 --> 01:08:42.260
shows you the true positive rate and
01:08:42.260 --> 01:08:43.990
the false positive rate for different
01:08:43.990 --> 01:08:44.700
thresholds.
01:08:45.460 --> 01:08:48.710
So I can choose a value such that L
01:08:48.710 --> 01:08:50.170
never have a.
01:08:50.940 --> 01:08:52.910
Where here I define true positive as y
01:08:52.910 --> 01:08:53.510
= 0.
01:08:54.220 --> 01:08:56.190
So I can choose a threshold where.
01:08:57.010 --> 01:08:59.930
I will get every single malign case
01:08:59.930 --> 01:09:02.380
correct, but I'll have like 20% false
01:09:02.380 --> 01:09:03.450
positives.
01:09:03.450 --> 01:09:05.870
Or I can choose a case where I'll
01:09:05.870 --> 01:09:07.360
sometimes make mistakes.
01:09:07.360 --> 01:09:10.110
Thinking I'm malignant is not
01:09:10.110 --> 01:09:11.040
malignant.
01:09:11.040 --> 01:09:15.360
But when it's benign, like 9099% of the
01:09:15.360 --> 01:09:16.570
time I'll think it's benign.
01:09:16.570 --> 01:09:18.815
So you can choose like you can kind of
01:09:18.815 --> 01:09:19.450
choose your errors.
01:09:25.800 --> 01:09:30.690
So this is so this like given some
01:09:30.690 --> 01:09:33.080
point on this curve, it tells me the
01:09:33.080 --> 01:09:35.120
true positive rate is the percent of
01:09:35.120 --> 01:09:37.775
times that I correctly classify equals
01:09:37.775 --> 01:09:39.379
zero as y = 0.
01:09:40.330 --> 01:09:42.020
And the false positive rate is the
01:09:42.020 --> 01:09:43.660
percent of times that I.
01:09:45.460 --> 01:09:46.790
Classify.
01:09:48.160 --> 01:09:50.400
Y = 1 as y = 0.
01:09:54.870 --> 01:09:57.410
Alright, so I can also look at the
01:09:57.410 --> 01:09:58.350
feature importance.
01:09:58.350 --> 01:10:01.450
So if I do L1, so here I trained one
01:10:01.450 --> 01:10:04.230
model with L1 logistic regression or
01:10:04.230 --> 01:10:06.586
this is L2 and one with L1 logistic
01:10:06.586 --> 01:10:06.930
regression?
01:10:07.740 --> 01:10:08.860
And that makes me use a different
01:10:08.860 --> 01:10:10.000
solver if it's L1.
01:10:11.270 --> 01:10:13.980
So I can see the errors.
01:10:14.070 --> 01:10:14.730
01:10:18.090 --> 01:10:19.505
A little weird but that error.
01:10:19.505 --> 01:10:24.588
But OK, I can see the errors and I can
01:10:24.588 --> 01:10:26.780
see the feature values.
01:10:29.290 --> 01:10:32.870
So with L2 I get lots of low weights,
01:10:32.870 --> 01:10:34.222
but none of them are zero.
01:10:34.222 --> 01:10:37.750
With L1 I get lots of 0 weights in a
01:10:37.750 --> 01:10:39.160
few larger weights.
01:10:43.420 --> 01:10:44.910
And then I can also do some further
01:10:44.910 --> 01:10:46.400
analysis looking at the tree.
01:10:48.090 --> 01:10:50.090
So first I'll train a full tree.
01:10:51.060 --> 01:10:53.010
And then next I'll train a tree with
01:10:53.010 --> 01:10:54.370
Max depth equals 2.
01:10:56.680 --> 01:11:00.006
So with the full tree I got error of
01:11:00.006 --> 01:11:00.403
4%.
01:11:00.403 --> 01:11:05.106
So it was as good as the OR was not as
01:11:05.106 --> 01:11:06.590
good as logistic regressor but pretty
01:11:06.590 --> 01:11:06.930
decent.
01:11:08.220 --> 01:11:09.500
But this tree is kind of hard to
01:11:09.500 --> 01:11:09.940
interpret.
01:11:09.940 --> 01:11:11.410
You wouldn't be able to give it to a
01:11:11.410 --> 01:11:13.415
technician and say like use this tree
01:11:13.415 --> 01:11:14.330
to make your decision.
01:11:15.050 --> 01:11:17.020
The short tree had higher error, but
01:11:17.020 --> 01:11:18.730
it's a lot simpler, so I can see its
01:11:18.730 --> 01:11:20.530
first splitting on the perimeter of the
01:11:20.530 --> 01:11:21.240
largest cells.
01:11:25.000 --> 01:11:27.510
And then finally, after doing all this
01:11:27.510 --> 01:11:30.010
analysis, I'm going to do tenfold cross
01:11:30.010 --> 01:11:32.780
validation using my best model.
01:11:33.370 --> 01:11:35.590
So here I'll just compare L1 logistic
01:11:35.590 --> 01:11:38.240
regression and nearest neighbor.
01:11:39.160 --> 01:11:41.345
I am doing tenfold, so I'm going to do
01:11:41.345 --> 01:11:45.126
10 estimates I do for each split.
01:11:45.126 --> 01:11:48.490
So the split will be after permutation.
01:11:48.490 --> 01:11:53.120
The first split will take indices 01020
01:11:53.120 --> 01:11:56.414
or yeah, 0102030, et cetera.
01:11:56.414 --> 01:12:00.540
The second split will take 11121, the
01:12:00.540 --> 01:12:03.840
third will take 21222, et cetera.
01:12:04.830 --> 01:12:07.050
Every time I use 90% of the data to
01:12:07.050 --> 01:12:09.400
train and the remaining data to test.
01:12:10.520 --> 01:12:12.510
And I'm doing that by just specifying
01:12:12.510 --> 01:12:13.990
the data that I'm using to test and
01:12:13.990 --> 01:12:15.930
then subtracting those indices to get
01:12:15.930 --> 01:12:17.100
the data that I used to train.
01:12:18.080 --> 01:12:21.396
Every time I normalize based on the
01:12:21.396 --> 01:12:23.140
training data, normalize both my
01:12:23.140 --> 01:12:24.554
training and validation data based on
01:12:24.554 --> 01:12:26.180
the same training data for the current
01:12:26.180 --> 01:12:26.540
split.
01:12:27.600 --> 01:12:29.340
Then I train and evaluate my nearest
01:12:29.340 --> 01:12:31.870
neighbor and logistic regressor.
01:12:38.000 --> 01:12:39.230
So that was fast.
01:12:40.850 --> 01:12:41.103
Right.
01:12:41.103 --> 01:12:43.950
And so then I have my errors.
01:12:43.950 --> 01:12:46.970
So one thing to note is that my even
01:12:46.970 --> 01:12:48.250
though in that one case I was
01:12:48.250 --> 01:12:50.310
evaluating before that one split, my
01:12:50.310 --> 01:12:52.190
logistic regression error was zero,
01:12:52.190 --> 01:12:53.670
it's not 0 every time.
01:12:53.670 --> 01:12:56.984
It ranges from zero to 5.3.
01:12:56.984 --> 01:12:59.906
And my nearest neighbor accuracy ranges
01:12:59.906 --> 01:13:02.980
from zero to 8 or 8.7 depending on the
01:13:02.980 --> 01:13:03.330
split.
01:13:04.300 --> 01:13:06.085
So different samples of your training
01:13:06.085 --> 01:13:08.592
and test data will give you different
01:13:08.592 --> 01:13:09.866
error measurement errors.
01:13:09.866 --> 01:13:11.950
And so that's why like cross validation
01:13:11.950 --> 01:13:14.300
can be a nice tool to give you not only
01:13:14.300 --> 01:13:16.870
an expected error, but some variance on
01:13:16.870 --> 01:13:18.140
the estimate of that error.
01:13:19.000 --> 01:13:19.500
So.
01:13:20.410 --> 01:13:23.330
My standard error of my estimate of the
01:13:23.330 --> 01:13:26.195
mean, which is the stair deviation of
01:13:26.195 --> 01:13:28.390
my error estimates divided by the
01:13:28.390 --> 01:13:29.720
square of the number of samples.
01:13:30.680 --> 01:13:35.420
Is 09 for nearest neighbor and six for
01:13:35.420 --> 01:13:36.500
logistic regression.
01:13:37.500 --> 01:13:39.270
And I can also use that to compute a
01:13:39.270 --> 01:13:41.540
confidence interval by multiplying that
01:13:41.540 --> 01:13:45.410
standard error by I forgot 1.96.
01:13:46.280 --> 01:13:49.330
So I can say like I'm 95% confident
01:13:49.330 --> 01:13:51.930
that my logistic regression error is
01:13:51.930 --> 01:13:56.440
somewhere between 12 and 34 or three.
01:13:56.440 --> 01:14:00.040
Sorry, 1.2% and 34%.
01:14:02.360 --> 01:14:04.615
And my nearest neighbor error is higher
01:14:04.615 --> 01:14:06.620
and I have like a bigger confidence
01:14:06.620 --> 01:14:07.020
interval.
01:14:09.360 --> 01:14:14.360
Now let's just compare very briefly how
01:14:14.360 --> 01:14:14.860
that.
01:14:15.610 --> 01:14:19.660
How the original paper did on this same
01:14:19.660 --> 01:14:20.110
problem?
01:14:23.320 --> 01:14:25.480
I just have one more slide, so don't
01:14:25.480 --> 01:14:27.950
worry, we will finish.
01:14:28.690 --> 01:14:30.360
Within a minute or so of runtime.
01:14:31.200 --> 01:14:33.610
Alright, so in the paper they use an
01:14:33.610 --> 01:14:36.300
MSM tree, which is that you have a
01:14:36.300 --> 01:14:37.820
linear classifier.
01:14:37.820 --> 01:14:39.240
Essentially that's used to do each
01:14:39.240 --> 01:14:40.140
split of the tree.
01:14:41.090 --> 01:14:42.720
But at the end of the day they choose
01:14:42.720 --> 01:14:44.550
only one split, so it ends up being a
01:14:44.550 --> 01:14:45.380
linear classifier.
01:14:46.300 --> 01:14:49.633
There they are trying to minimize the
01:14:49.633 --> 01:14:51.520
number of features as well as the
01:14:51.520 --> 01:14:53.900
number of splitting planes in order to
01:14:53.900 --> 01:14:55.550
improve generalization and make a
01:14:55.550 --> 01:14:57.090
simple interpretable function.
01:14:57.800 --> 01:14:59.370
So at the end of the day, they choose
01:14:59.370 --> 01:15:01.105
just three features, mean texture,
01:15:01.105 --> 01:15:02.780
worst area and worst smoothness.
01:15:03.520 --> 01:15:04.420
And.
01:15:05.930 --> 01:15:08.610
They used tenfold cross validation and
01:15:08.610 --> 01:15:11.770
they got an error of 3% within a
01:15:11.770 --> 01:15:15.570
confidence interval or minus 15%.
01:15:15.570 --> 01:15:17.120
So pretty similar to what we got.
01:15:17.120 --> 01:15:18.960
We got slightly lower error but we were
01:15:18.960 --> 01:15:20.560
using more features in the logistic
01:15:20.560 --> 01:15:21.090
regressor.
01:15:21.910 --> 01:15:23.694
And then they tested it on their held
01:15:23.694 --> 01:15:26.475
out set and they got a perfect accuracy
01:15:26.475 --> 01:15:27.730
on the held out set.
01:15:28.550 --> 01:15:29.849
Now that doesn't mean that their
01:15:29.850 --> 01:15:31.670
accuracy is perfect because they're
01:15:31.670 --> 01:15:34.350
cross validation if anything, is a
01:15:34.350 --> 01:15:37.315
biased towards a underestimating the
01:15:37.315 --> 01:15:37.570
error.
01:15:37.570 --> 01:15:40.440
So I would say their error is like
01:15:40.440 --> 01:15:43.870
roughly 15 to 45%, which is what they
01:15:43.870 --> 01:15:45.180
correctly report in the paper.
01:15:46.950 --> 01:15:47.290
Right.
01:15:47.290 --> 01:15:51.030
So we performed fairly similarly to the
01:15:51.030 --> 01:15:51.705
analysis.
01:15:51.705 --> 01:15:53.670
The nice thing is that now I can do
01:15:53.670 --> 01:15:56.900
this like in under an hour if I want.
01:15:56.900 --> 01:15:59.140
Well at that time it would be a lot
01:15:59.140 --> 01:16:01.380
more work to do that kind of analysis.
01:16:02.330 --> 01:16:04.490
But they also need to obviously want to
01:16:04.490 --> 01:16:06.410
be a lot more careful and do careful
01:16:06.410 --> 01:16:07.780
analysis and make sure that this is
01:16:07.780 --> 01:16:10.240
going to be like a useful tool for.
01:16:10.320 --> 01:16:12.180
That guy's diagnosis.
01:16:14.130 --> 01:16:14.870
Hey.
01:16:14.870 --> 01:16:16.400
So hopefully that was helpful.
01:16:16.400 --> 01:16:19.700
And next week I am going to talk about
01:16:19.700 --> 01:16:20.150
or not.
01:16:20.150 --> 01:16:21.750
Next week it's only Tuesday.
01:16:21.750 --> 01:16:23.550
On Thursday I'm going to talk about.
01:16:23.550 --> 01:16:24.962
No, wait, what day is it?
01:16:24.962 --> 01:16:25.250
Thursday.
01:16:25.250 --> 01:16:25.868
OK, good.
01:16:25.868 --> 01:16:27.020
It is next week.
01:16:27.020 --> 01:16:28.810
Yeah, at least chat with time.
01:16:30.520 --> 01:16:33.300
Next week I'll talk about ensembles and
01:16:33.300 --> 01:16:35.310
SVM and stochastic gradient descent.
01:16:35.310 --> 01:16:35.780
Thanks.
01:16:35.780 --> 01:16:36.690
Have a good weekend.
01:16:38.360 --> 01:16:40.130
And remember that homework one is due
01:16:40.130 --> 01:16:40.830
Monday.
01:16:41.650 --> 01:16:42.760
For those asking question.