Would you like to inspect the original subtitles? These are the user uploaded subtitles that are being translated:
0
00:00:00,940 --> 00:00:01,590
[Autogenerated] when you're building
1
00:00:01,590 --> 00:00:03,270
machine learning models, it's common
2
00:00:03,270 --> 00:00:05,139
practice to split up the data that you
3
00:00:05,139 --> 00:00:07,860
have to work with into two, maybe even
4
00:00:07,860 --> 00:00:10,099
three subsets training tests and
5
00:00:10,099 --> 00:00:12,769
validation data. Let's talk about why we
6
00:00:12,769 --> 00:00:15,419
do this and how this fits into the machine
7
00:00:15,419 --> 00:00:17,609
learning workflow. We know the basic ___
8
00:00:17,609 --> 00:00:19,210
and board in getting your machine learning
9
00:00:19,210 --> 00:00:22,300
model to production. Observe the step here
10
00:00:22,300 --> 00:00:24,129
where you train your model. Choose a
11
00:00:24,129 --> 00:00:26,640
validation method. Validate your model,
12
00:00:26,640 --> 00:00:29,620
examine and score your model and it read
13
00:00:29,620 --> 00:00:31,570
this process till you're satisfied with
14
00:00:31,570 --> 00:00:33,780
how we're model. Performs when we talk
15
00:00:33,780 --> 00:00:35,859
off, splitting our data into subsets,
16
00:00:35,859 --> 00:00:38,000
these other steps that will use those
17
00:00:38,000 --> 00:00:40,359
subsets of data to validate and evaluate
18
00:00:40,359 --> 00:00:43,250
our model. Let's say this is all the data
19
00:00:43,250 --> 00:00:45,090
that we have available in the world to
20
00:00:45,090 --> 00:00:47,679
feed into our model. This is our entire
21
00:00:47,679 --> 00:00:50,259
world off data. Now let's say you want to
22
00:00:50,259 --> 00:00:52,460
take the view that you want as much data
23
00:00:52,460 --> 00:00:55,049
as possible to train your model. You'll
24
00:00:55,049 --> 00:00:57,079
take all of the state of that. You have
25
00:00:57,079 --> 00:00:59,609
pre process it in the right way and feed
26
00:00:59,609 --> 00:01:01,630
it into building a machine learning model.
27
00:01:01,630 --> 00:01:04,290
But there's a problem with that data that
28
00:01:04,290 --> 00:01:06,859
you used to train a model cannot be used
29
00:01:06,859 --> 00:01:08,890
to evaluate a model. Once you've trained
30
00:01:08,890 --> 00:01:10,010
your mortal, you want to know where that
31
00:01:10,010 --> 00:01:12,349
it's a good one. You can't use the same
32
00:01:12,349 --> 00:01:14,650
training data for that. That's because of
33
00:01:14,650 --> 00:01:17,069
the model has seen all of the instances in
34
00:01:17,069 --> 00:01:19,379
the training data during the process of
35
00:01:19,379 --> 00:01:22,370
trading. So it's quite possible that if
36
00:01:22,370 --> 00:01:23,939
you evaluate your model on the training
37
00:01:23,939 --> 00:01:26,739
data, the model gives you a great score.
38
00:01:26,739 --> 00:01:28,500
But that may not mean that it's a good
39
00:01:28,500 --> 00:01:30,680
model. The model we have memorized
40
00:01:30,680 --> 00:01:32,900
training instances. And if you want to,
41
00:01:32,900 --> 00:01:35,230
indeed uses model on instances that it
42
00:01:35,230 --> 00:01:37,519
hasn't encountered before, it would
43
00:01:37,519 --> 00:01:40,379
perform poorly. It may have over fit on
44
00:01:40,379 --> 00:01:43,140
the training data, so mortal robustness
45
00:01:43,140 --> 00:01:45,730
cannot be measured on instances the model
46
00:01:45,730 --> 00:01:48,519
has encountered before. So if you have all
47
00:01:48,519 --> 00:01:50,049
of the state and the ball and you use
48
00:01:50,049 --> 00:01:52,569
everything to train your model, well, then
49
00:01:52,569 --> 00:01:54,260
you're left with nothing to evaluate your
50
00:01:54,260 --> 00:01:57,209
model. So what do you do now? So you have
51
00:01:57,209 --> 00:01:59,319
all of this data available to you. You
52
00:01:59,319 --> 00:02:02,269
split it into two subsets training data
53
00:02:02,269 --> 00:02:05,810
and test data. It's quite common to use
54
00:02:05,810 --> 00:02:08,699
80% off the data to train the model. This
55
00:02:08,699 --> 00:02:10,889
is referred to as training data, and this
56
00:02:10,889 --> 00:02:12,780
is the only set of instances that your
57
00:02:12,780 --> 00:02:15,550
model we see in the training process you
58
00:02:15,550 --> 00:02:17,860
set aside a portion of the original data.
59
00:02:17,860 --> 00:02:21,819
Let's it's 20% to sanity. Check or measure
60
00:02:21,819 --> 00:02:24,159
the performance off your Marty during the
61
00:02:24,159 --> 00:02:26,129
training process. Your model will never
62
00:02:26,129 --> 00:02:29,139
encounter the test data, which means if
63
00:02:29,139 --> 00:02:31,009
you want to evaluate whether your model is
64
00:02:31,009 --> 00:02:32,789
a robust one, that it works well on
65
00:02:32,789 --> 00:02:35,340
instances it hasn't seen before. Well, you
66
00:02:35,340 --> 00:02:37,780
lose the test data off of that. You will
67
00:02:37,780 --> 00:02:39,740
run through one training process to
68
00:02:39,740 --> 00:02:43,189
generate one candidate model. You'll set
69
00:02:43,189 --> 00:02:45,400
up a model with a certain design, and
70
00:02:45,400 --> 00:02:47,860
you'll get one candidate model at the end
71
00:02:47,860 --> 00:02:50,189
of the training process. If you want any
72
00:02:50,189 --> 00:02:52,349
different candidate models, you'll need to
73
00:02:52,349 --> 00:02:55,400
run in different training processes. And
74
00:02:55,400 --> 00:02:57,560
for each of these end candidate models,
75
00:02:57,560 --> 00:03:00,000
you'll run and test processes toe.
76
00:03:00,000 --> 00:03:02,240
Evaluate these models and then you'll pick
77
00:03:02,240 --> 00:03:04,949
the best one. So far, this is seeming like
78
00:03:04,949 --> 00:03:07,139
a pretty good idea. You have training data
79
00:03:07,139 --> 00:03:10,150
and you have test data. The test data are
80
00:03:10,150 --> 00:03:12,139
the tests that can be used to choose the
81
00:03:12,139 --> 00:03:15,009
best candidate model, which means you're
82
00:03:15,009 --> 00:03:17,719
evaluating our modern instances the model
83
00:03:17,719 --> 00:03:20,449
hasn't seen before. It hasn't seen during
84
00:03:20,449 --> 00:03:22,669
the training process, but there's a
85
00:03:22,669 --> 00:03:25,090
problem here as well. It's quite possible
86
00:03:25,090 --> 00:03:27,550
that your evaluation itself can become
87
00:03:27,550 --> 00:03:30,240
bias when you use the same set of
88
00:03:30,240 --> 00:03:32,650
instances that is the test set toe pick
89
00:03:32,650 --> 00:03:35,139
the best Candidate model. This can lead to
90
00:03:35,139 --> 00:03:37,500
a kind off over fitting. This is refer to
91
00:03:37,500 --> 00:03:40,400
us over fitting on the test set. When you
92
00:03:40,400 --> 00:03:42,430
have data split into just two sets
93
00:03:42,430 --> 00:03:44,990
training and test your candidate model can
94
00:03:44,990 --> 00:03:47,560
end up or fitting on the test it. And once
95
00:03:47,560 --> 00:03:49,689
again, you haven't really got a robust
96
00:03:49,689 --> 00:03:52,759
model her, which leads us to the next
97
00:03:52,759 --> 00:03:55,789
option. Cross validation. Instead of using
98
00:03:55,789 --> 00:03:58,340
just one test set to evaluate your model,
99
00:03:58,340 --> 00:04:01,050
you'll come out a separate validation set
100
00:04:01,050 --> 00:04:03,240
of data points. Now you'll generate
101
00:04:03,240 --> 00:04:05,310
different candidate models using the
102
00:04:05,310 --> 00:04:07,300
training set. You'll then use the
103
00:04:07,300 --> 00:04:09,879
validation set to pick the best candidate
104
00:04:09,879 --> 00:04:12,530
model, and then the final evaluation off.
105
00:04:12,530 --> 00:04:15,240
That model will be on the test set. Let's
106
00:04:15,240 --> 00:04:16,899
take a look at how this works visually,
107
00:04:16,899 --> 00:04:18,569
you have all of the data that you're
108
00:04:18,569 --> 00:04:20,790
working with in the world, you split it
109
00:04:20,790 --> 00:04:23,319
into three subsets. Most of this will be
110
00:04:23,319 --> 00:04:25,060
training data that you used to train the
111
00:04:25,060 --> 00:04:27,290
model. You will have two other subsets
112
00:04:27,290 --> 00:04:30,779
validation data and test data, so you're
113
00:04:30,779 --> 00:04:33,930
holding out to subset from the original
114
00:04:33,930 --> 00:04:36,829
creaming process. So let's say you want to
115
00:04:36,829 --> 00:04:38,829
build a number of different models and
116
00:04:38,829 --> 00:04:40,930
then choose the best one. You lose the
117
00:04:40,930 --> 00:04:43,189
training data to produce the different
118
00:04:43,189 --> 00:04:45,170
candidate models in the news. The
119
00:04:45,170 --> 00:04:47,819
validation data toe. Evaluate all of these
120
00:04:47,819 --> 00:04:50,709
models once you figured out which model is
121
00:04:50,709 --> 00:04:52,810
the best one. Based on the validation
122
00:04:52,810 --> 00:04:55,970
data, you let finally evaluate that
123
00:04:55,970 --> 00:04:59,269
candidate model on the test data. The test
124
00:04:59,269 --> 00:05:01,009
data is comprised of instances that your
125
00:05:01,009 --> 00:05:04,000
model has never seen before. Either during
126
00:05:04,000 --> 00:05:07,250
training are during validation. Test data
127
00:05:07,250 --> 00:05:09,589
gives you an unbiased evaluation off your
128
00:05:09,589 --> 00:05:12,600
model, and with this set up, you can see
129
00:05:12,600 --> 00:05:14,639
whether your finally model is a robust
130
00:05:14,639 --> 00:05:16,660
one. You can also generate multiple
131
00:05:16,660 --> 00:05:19,230
candidate models and select the best one.
132
00:05:19,230 --> 00:05:21,490
This process is called hyper parameter
133
00:05:21,490 --> 00:05:23,920
tuning, and you can evaluate the final
134
00:05:23,920 --> 00:05:26,800
candidate model on your test data so
135
00:05:26,800 --> 00:05:28,750
you'll run and training process is to
136
00:05:28,750 --> 00:05:35,000
generate end candidate models n validation processes and one test process
10773
Can't find what you're looking for?
Get subtitles in any language from opensubtitles.com, and translate them here.