1
00:00:11,100 --> 00:00:16,770
In this video, we are going to summarize everything we learned in the section, this section focused

2
00:00:16,770 --> 00:00:19,000
on convolutional neural networks.

3
00:00:19,440 --> 00:00:25,560
Now for us, since we already knew how a Anan's work, the only matter at hand was to understand convolution

4
00:00:25,740 --> 00:00:27,600
and why it's appropriate in deep learning.

5
00:00:28,710 --> 00:00:33,810
I did give you the option to skip this part if you are a beginner at math and just wanted to see the

6
00:00:33,810 --> 00:00:39,750
code so only those who watch the convolution lectures will have a good understanding of how it works

7
00:00:39,750 --> 00:00:40,920
and why we use it.

8
00:00:42,150 --> 00:00:45,690
These videos taught you convolution from multiple different perspectives.

9
00:00:46,020 --> 00:00:51,960
I think this is useful because people often believe convolution is abstract, but in fact it's very

10
00:00:51,960 --> 00:00:54,990
intuitive from the perspective of image processing.

11
00:00:56,280 --> 00:00:58,710
We first looked at how to perform convolution.

12
00:00:59,130 --> 00:01:03,690
That is what convolution actually does in terms of multiplication and addition.

13
00:01:04,770 --> 00:01:07,830
Surprisingly, all it takes is a bit of 3rd grade math.

14
00:01:08,220 --> 00:01:11,430
Now, understanding what that does takes a more mature perspective.

15
00:01:11,760 --> 00:01:13,860
So it's not just about the mechanics.

16
00:01:15,210 --> 00:01:20,030
We learn that convolution and deep learning is really what everyone else calls cross correlation.

17
00:01:20,550 --> 00:01:25,950
And in fact, correlation is probably a much better name, since most people have a better understanding

18
00:01:26,100 --> 00:01:27,570
of what this word means.

19
00:01:28,110 --> 00:01:34,940
Correlation is really just two words put together correlation, the relationship between two things.

20
00:01:35,370 --> 00:01:38,240
Thus it tells us how related two things are.

21
00:01:38,430 --> 00:01:39,090
Time series.

22
00:01:39,090 --> 00:01:44,220
In our case, this leads us to the idea that convolution is a pattern matcher.

23
00:01:44,760 --> 00:01:50,250
The convolution filter is like a template that we want to match with and the input image is the thing

24
00:01:50,250 --> 00:01:52,500
where we want to find relevant patterns.

25
00:01:57,170 --> 00:02:03,290
We also noted that the Arima models we learned about are really just convolutions, a Remar Perimeter's

26
00:02:03,290 --> 00:02:06,440
or Arima weights are really just a convolution filter.

27
00:02:07,070 --> 00:02:11,390
Thus we are already done convolution in this course without knowing we did.

28
00:02:16,100 --> 00:02:22,310
After gaining a deeper understanding about convolution, we then learned how to build Xians, we observe

29
00:02:22,310 --> 00:02:26,440
that after learning how to build Anan's building, CNN's is pretty easy.

30
00:02:26,690 --> 00:02:29,340
Just a few new layers, as per usual.

31
00:02:29,360 --> 00:02:32,000
We then applied science to several data sets.

32
00:02:33,380 --> 00:02:38,570
At this point, let's remind ourselves that there are so many combinations of things to try in this

33
00:02:38,570 --> 00:02:39,190
cause.

34
00:02:39,680 --> 00:02:42,400
For example, there are other data sets you could have tried.

35
00:02:42,740 --> 00:02:48,560
You can add more layers, you can take away layers, you can use flatten instead of global max pooling.

36
00:02:49,220 --> 00:02:53,810
You can use technique activations instead of real use for stock returns.

37
00:02:53,810 --> 00:02:57,900
You can add different features, for example, those based on technical indicators.

38
00:02:58,340 --> 00:03:00,630
So it really the possibilities are endless.

39
00:03:01,010 --> 00:03:06,140
You've been taught all the code you need to know, but if you have any ideas, you want to try using

40
00:03:06,140 --> 00:03:09,530
that code, please do so and share your results on the Q&A.