1
00:00:11,040 --> 00:00:16,920
In this video, we will be introduced to the next section of this course, the main topic of this section

2
00:00:16,920 --> 00:00:18,960
is convolutional neural networks.

3
00:00:19,620 --> 00:00:21,380
We also call these CNN.

4
00:00:21,990 --> 00:00:25,700
So basically a rough outline of this section is just two steps.

5
00:00:26,160 --> 00:00:29,720
Number one, explain what a CNN is and how it works.

6
00:00:29,940 --> 00:00:36,990
And number to show you how to use CNN to forecast and classify time series using various data sets as

7
00:00:36,990 --> 00:00:37,620
per usual.

8
00:00:38,730 --> 00:00:44,010
One interesting thing to look out for is whether or not CNN will be more powerful than what we learned

9
00:00:44,010 --> 00:00:45,810
about in the previous sections.

10
00:00:50,380 --> 00:00:57,220
OK, so as mentioned, the first part of this section will focus on what CNN is now, this may be surprising,

11
00:00:57,220 --> 00:00:59,050
but the name actually tells you quite a bit.

12
00:00:59,560 --> 00:01:04,180
A convolutional neural network is really just a neural network with convolutions.

13
00:01:04,720 --> 00:01:09,820
Thus, a major component of this section will be to discuss what is convolution.

14
00:01:10,570 --> 00:01:12,250
You already know the neural network part.

15
00:01:13,210 --> 00:01:16,190
Therefore, all that remains is the convolution part.

16
00:01:16,510 --> 00:01:18,460
And of course, how to put them together.

17
00:01:19,720 --> 00:01:25,030
In my opinion, the most intuitive way to understand convolution is by understanding how they can be

18
00:01:25,030 --> 00:01:26,080
used for images.

19
00:01:26,350 --> 00:01:32,660
That is, pictures, pictures like Time series or objects that we can see on our computer screens.

20
00:01:33,010 --> 00:01:38,740
So it is obvious when we do something to an image what kind of effect it has simply because we can see

21
00:01:38,740 --> 00:01:38,930
it.

22
00:01:39,700 --> 00:01:44,050
Now, this is the case for Time series, too, but Times series are a bit more abstract.

23
00:01:44,530 --> 00:01:50,620
Because of this, we will first study convolution in the context of images after doing this and gaining

24
00:01:50,620 --> 00:01:53,330
some intuition on what convolution actually does.

25
00:01:53,620 --> 00:01:55,900
We will learn how they are applied to Time series.

26
00:01:56,470 --> 00:02:02,650
Mathematically, convolution for a Time series is simpler than for images, but intuitively images are

27
00:02:02,650 --> 00:02:03,980
simpler than Time series.

28
00:02:04,390 --> 00:02:07,690
So there's a bit of a tradeoff there, depending on how you see things.

29
00:02:08,050 --> 00:02:13,030
When you look at the math time series are simpler, but when you look at the intuition, images are

30
00:02:13,030 --> 00:02:13,570
simpler.

31
00:02:14,440 --> 00:02:20,380
Do note, however, that time series can be converted into images and then regular image convolutions

32
00:02:20,380 --> 00:02:21,490
can be used.

33
00:02:25,960 --> 00:02:30,820
Now, I want to provide a bit of warning to those beginner students who might be intimidated by math

34
00:02:31,270 --> 00:02:37,610
convolution is a mathematical operation, however intenser flow, it's just one line of code.

35
00:02:38,050 --> 00:02:43,540
So if you are intimidated by math, keep in mind that you can potentially just skip the explanation

36
00:02:43,540 --> 00:02:49,390
for what convolution is and simply treat the code as if you are using APIs, which is in fact what is

37
00:02:49,390 --> 00:02:50,080
being done.

38
00:02:50,800 --> 00:02:55,950
If you decide to go this route, then you can skip to the lecture called CNN Code Preparation.

39
00:02:56,440 --> 00:03:01,420
But if you do want to understand convolution and why they are useful in deep learning, then you should

40
00:03:01,420 --> 00:03:02,650
watch all the lectures.