1
00:00:11,080 --> 00:00:17,260
In this video, we will be introduced to the next section of this course, the main topic of this section

2
00:00:17,260 --> 00:00:18,920
is recurrent neural networks.

3
00:00:19,420 --> 00:00:21,110
We also call these Arnon's.

4
00:00:21,880 --> 00:00:23,420
So what makes art and special?

5
00:00:24,250 --> 00:00:26,890
Well, consider the neural networks we have studied so far.

6
00:00:27,310 --> 00:00:32,640
They all of this one peculiar property, which makes them unlike biological neural networks.

7
00:00:33,250 --> 00:00:38,320
And this is that all the information goes in the same direction from left to right.

8
00:00:39,100 --> 00:00:40,920
So here's a question to consider.

9
00:00:41,290 --> 00:00:46,660
Why not simply have neuronal connections which are less restrictive, for example, ones that can go

10
00:00:46,660 --> 00:00:52,940
backwards when we create connections that go backwards in a specific way with a time delay of one step?

11
00:00:53,200 --> 00:00:55,340
We call these recurrent neural networks.

12
00:00:56,230 --> 00:01:01,120
Now, please note that there are all kinds of ways we could make new connections, but only this specific

13
00:01:01,120 --> 00:01:03,580
way is what we call a recurrent neural network.

14
00:01:04,540 --> 00:01:09,610
OK, so this diagram is pretty messy, but what you should notice is that every hearing unit has an

15
00:01:09,610 --> 00:01:13,840
arrow that goes back towards all other hidden units, including itself.

16
00:01:14,620 --> 00:01:20,200
The most common diagram, which looks like this is a bit cleaner but actually does not reflect how the

17
00:01:20,200 --> 00:01:22,690
connections are made in this case.

18
00:01:22,720 --> 00:01:26,050
Notice how each hitting unit loops back only to itself.

19
00:01:26,410 --> 00:01:28,180
This is not exactly what we want.

20
00:01:33,020 --> 00:01:38,720
So let's do a brief outline of this section, as per usual, this section will start with the theory

21
00:01:38,720 --> 00:01:40,980
and concepts behind Arnett's.

22
00:01:41,390 --> 00:01:44,770
Well, first study, the simple Arnim, also known as the elimination.

23
00:01:44,770 --> 00:01:51,050
And this will form the basis for everything else we do once we understand the simple answer.

24
00:01:51,050 --> 00:01:54,770
And we'll do a small code exercise which will reinforce your understanding.

25
00:01:55,850 --> 00:02:00,350
As you recall, one of the major themes of this course is thinking about shapes.

26
00:02:00,830 --> 00:02:03,890
This section will really test your ability to understand this.

27
00:02:04,610 --> 00:02:09,110
Now, when it comes to Arnon's, the simple answer is rarely ever used any more.

28
00:02:09,950 --> 00:02:13,930
For the time being, the LTM is the most common type of Afnan unit.

29
00:02:14,420 --> 00:02:18,530
So we'll study the LSM and the closely related Ajamu.

30
00:02:19,040 --> 00:02:24,020
You'll see why the l'Est Yemanja you are needed and what advantages they have over.

31
00:02:24,020 --> 00:02:30,710
The simple answer in the next step will be to put everything into practice with various datasets as

32
00:02:30,710 --> 00:02:31,220
usual.

33
00:02:32,120 --> 00:02:36,370
Now I want you to remember one thing, which is that all data is the same.

34
00:02:36,830 --> 00:02:42,050
So although Arnon's can be applied to all of the data sets we've looked at so far, we of course do

35
00:02:42,050 --> 00:02:43,300
not have time for the model.

36
00:02:43,790 --> 00:02:48,890
So if there's anything you want to try them on, please remember that no code has to change.

37
00:02:49,190 --> 00:02:51,740
Simply switch one data set out for another.

38
00:02:52,220 --> 00:02:56,680
And of course, there will be plenty of other configuration options to choose from as well.

39
00:02:57,080 --> 00:02:59,780
So please feel free to fiddle around and test things out.

40
00:03:04,530 --> 00:03:09,240
Now, I want to make a special note that, like with convolution, you don't have to go through the

41
00:03:09,240 --> 00:03:12,590
three lectures if you find the math is intimidating for you.

42
00:03:13,020 --> 00:03:18,070
Instead, if you prefer to treat the code as simply like an API, you can do that as well.

43
00:03:18,630 --> 00:03:22,770
So if that's the strategy you wish to take, please skip to the code preparation.

44
00:03:24,180 --> 00:03:29,680
In my opinion, Arnon's contain a bit less math and science, so they might be easier to understand.

45
00:03:30,240 --> 00:03:33,850
It's mostly about understanding the diagrams and the flow of the data.

46
00:03:34,500 --> 00:03:38,370
However, understanding the equations you see will make it much simpler.

47
00:03:39,270 --> 00:03:44,520
This actually goes back to our discussion on machine learning, where I said that one of the most important

48
00:03:44,520 --> 00:03:49,800
expressions you will learn is W, transpose X plus B for Arnon's.

49
00:03:49,800 --> 00:03:52,320
We are going to see this in multiple places at once.

50
00:03:52,920 --> 00:03:58,350
Recall that I advised you not to think about this as math, but merely a pattern or an operation.

51
00:03:58,950 --> 00:04:04,200
So once you see things from this perspective, it makes it easier since you can think at a higher level

52
00:04:04,350 --> 00:04:09,450
about how these more fundamental operations can be connected together to make new things.

53
00:04:10,080 --> 00:04:15,360
In fact, I encourage you to watch this section with the perspective that there actually is no math

54
00:04:15,360 --> 00:04:19,470
at all, although you're going to see some equations that look like math.

55
00:04:19,620 --> 00:04:23,870
It's actually very interesting to note that you don't have to actually do any math.

56
00:04:24,390 --> 00:04:28,800
The section is more about understanding things at a higher level and how different building blocks fit

57
00:04:28,800 --> 00:04:31,060
together to create bigger systems.

58
00:04:31,620 --> 00:04:37,200
Of course, it's possible you might be so intimidated by the math that you'll see it in all this advice

59
00:04:37,200 --> 00:04:40,880
will go out the window so you can make that determination for yourself.