1
00:00:11,080 --> 00:00:16,149
So in this lecture, we'll be transitioning into the next part of the course, which is on deep learning

2
00:00:16,149 --> 00:00:17,440
and deep neural networks.

3
00:00:18,010 --> 00:00:20,680
The previous part on machine learning is now complete.

4
00:00:21,610 --> 00:00:23,590
The goal of this section is pretty simple.

5
00:00:23,740 --> 00:00:29,320
You'll learn about deep neural networks, which will allow us to remove many of the simplifying assumptions

6
00:00:29,320 --> 00:00:31,570
we used earlier in this course.

7
00:00:32,290 --> 00:00:38,470
For instance, instead of bag of words techniques like TF IDF, we'll use embeddings instead of the

8
00:00:38,470 --> 00:00:39,460
mark of assumption.

9
00:00:39,790 --> 00:00:45,520
We simply won't make any assumptions about how many passwords can influence the next word in a sentence.

10
00:00:46,090 --> 00:00:52,750
Instead of using a linear models like SVD to transform our data insulating factors, we can use nonlinear

11
00:00:52,750 --> 00:00:55,060
neural networks, which are much more powerful.

12
00:00:59,740 --> 00:01:02,260
So the outline for this module is very simple.

13
00:01:02,890 --> 00:01:08,500
We'll start with learning the basic syntax of TensorFlow and how it can be used to build machines that

14
00:01:08,500 --> 00:01:09,010
learn.

15
00:01:09,730 --> 00:01:14,950
We'll do this by first using linear models, which in the study of neural networks are a model of the

16
00:01:14,950 --> 00:01:15,550
neuron.

17
00:01:16,210 --> 00:01:21,070
If you don't have a background in biology, all you have to know is that the neuron is a cell in your

18
00:01:21,070 --> 00:01:27,190
brain, and you have many of these sitting inside your head which do things like process information

19
00:01:27,460 --> 00:01:30,310
and make decisions about what you should do with yourself.

20
00:01:30,880 --> 00:01:36,580
Basically, everything you feel, every thought you have and every action you take is because of your

21
00:01:36,580 --> 00:01:37,300
neurons.

22
00:01:37,900 --> 00:01:43,390
It turns out that this a neuron model is the same as logistic regression, which, as you may recall,

23
00:01:43,390 --> 00:01:44,710
we've already studied.

24
00:01:45,370 --> 00:01:51,070
So this will be a new perspective on that model and obviously a new way to implement one as well.

25
00:01:51,850 --> 00:01:57,310
Once we've learned about the basics of neurons and TensorFlow syntax, we'll move on to artificial neural

26
00:01:57,310 --> 00:01:58,090
networks.

27
00:01:58,660 --> 00:02:00,910
This is one we'll also learn about embeddings.

28
00:02:01,510 --> 00:02:06,760
Embeddings are a better method of putting text into neural networks compared to simple bag of words

29
00:02:06,760 --> 00:02:08,259
models like Tia Faria F.

30
00:02:09,759 --> 00:02:15,130
Once we've learned about Ian's will then learn about CNN's or convolutional neural networks.

31
00:02:15,610 --> 00:02:21,460
These were originally invented for images, but as you'll see, they can be used for any kind of continuous

32
00:02:21,460 --> 00:02:25,540
signal, which is what we get when we convert our text into embeddings.

33
00:02:26,140 --> 00:02:30,640
Basically, we learn how to convert text into a multi-dimensional time series.

34
00:02:31,360 --> 00:02:36,340
The third section will cover recurrent neural networks, or aren't ends, which are neural networks

35
00:02:36,340 --> 00:02:39,130
specifically designed with sequences in mind.

36
00:02:39,670 --> 00:02:44,170
We'll look at the most popular variants of our own ends, like the Elysium and the Gear you.

37
00:02:48,970 --> 00:02:54,280
So the final thing I want to mention in this lecture is the strategy you should take if you've already

38
00:02:54,280 --> 00:02:57,400
studied TensorFlow or deep learning with me in the past.

39
00:02:58,000 --> 00:03:00,820
Obviously, the theory behind deep learning hasn't changed.

40
00:03:01,210 --> 00:03:03,850
So there's nothing new to learn about in that regard.

41
00:03:04,360 --> 00:03:09,700
If you already know about Annan's CNN's inheritance, either from me or from somewhere else, it is

42
00:03:09,700 --> 00:03:14,460
safe to simply skip to the code preparation or even just the code lectures themselves.

43
00:03:16,720 --> 00:03:21,730
So you might also be wondering what's the difference between taking this course and of course, I've

44
00:03:21,730 --> 00:03:24,160
made which actually focuses on deep learning.

45
00:03:24,970 --> 00:03:29,330
The answer should make pretty logical sense in my deep learning courses.

46
00:03:29,350 --> 00:03:34,750
You learn more about various aspects of deep learning, like why and how neural networks work.

47
00:03:35,470 --> 00:03:40,960
One major part of this is the different types of laws functions in why we use them, and also the exact

48
00:03:40,960 --> 00:03:44,320
details of optimization methods like Adam A.G D..

49
00:03:45,520 --> 00:03:50,890
This course will not go over those because if you want to learn the intricate details about deep learning,

50
00:03:51,190 --> 00:03:53,500
well, then you should take a deep learning course.

51
00:03:54,190 --> 00:04:00,160
On the other hand, a generic deep learning course will not focus on NLP, which this course obviously

52
00:04:00,160 --> 00:04:00,790
does.

53
00:04:01,510 --> 00:04:07,420
So although you can apply neural networks to all kinds of things like disease prediction, fraud detection,

54
00:04:07,420 --> 00:04:11,920
images, time series and so forth, we will only apply them to NLP.

55
00:04:12,670 --> 00:04:17,980
So you can think of this like a Venn diagram where there's a little bit of crossover, but neither is

56
00:04:17,980 --> 00:04:19,360
a superset of the other.

57
00:04:20,079 --> 00:04:24,610
This course will focus on some parts of deep learning, while the deep learning course, which I've

58
00:04:24,610 --> 00:04:27,310
taught will only cover some of NLP.

59
00:04:28,060 --> 00:04:33,130
Essentially in this course, we are mostly concerned with how to apply deep learning to NLP.

60
00:04:34,000 --> 00:04:40,120
So an example of a topic that we should not care about is why the cross entropy loss is the right laws

61
00:04:40,120 --> 00:04:41,560
for classifying text.

62
00:04:42,250 --> 00:04:45,430
And this course, I'm simply going to tell you that this is the case.

63
00:04:45,850 --> 00:04:51,130
And if you want to study deep learning more in depth to learn about why I have left resources for you

64
00:04:51,130 --> 00:04:52,330
in the GitHub repo.