1
00:00:02,070 --> 00:00:10,170
Now we are going to begin our journey into the field of neural networks, work on neural networks,

2
00:00:10,530 --> 00:00:16,980
began with the motivation that human brain computes in a very different way than a conventional computer.

3
00:00:18,690 --> 00:00:25,620
And human brain is able to perform certain types of computations, such as image recognition and pattern

4
00:00:25,620 --> 00:00:28,890
recognition, much better than a conventional computer.

5
00:00:30,360 --> 00:00:37,500
On the other hand, computers are very good at complex arithmetic calculations, which human brain is

6
00:00:37,590 --> 00:00:38,460
not very good at.

7
00:00:41,700 --> 00:00:45,300
Let's take an example of recognizing handed Inderjit.

8
00:00:46,650 --> 00:00:55,230
Recognizing so many different types of handwritten digits is not easy, but we humans are astoundingly

9
00:00:55,230 --> 00:00:57,720
good at making sense of these digits.

10
00:00:59,160 --> 00:01:04,950
On the other hand, if we have to write a computer program to the recognize did it like these.

11
00:01:05,820 --> 00:01:07,530
Can you imagine how difficult it is?

12
00:01:08,730 --> 00:01:18,010
What will be the tool to see that these two are the same digit nine and this one is a seven and not

13
00:01:18,010 --> 00:01:18,450
a nine?

14
00:01:21,090 --> 00:01:28,680
When we try to make such precise rules, we get stuck with a number of exceptions, caveats and special

15
00:01:28,680 --> 00:01:29,220
cases.

16
00:01:31,020 --> 00:01:33,090
This is where neural networks come in.

17
00:01:35,280 --> 00:01:38,820
Neural networks do not have the inbuilt rules.

18
00:01:39,450 --> 00:01:48,360
Rather, neural networks infer these rules for recognizing handwriting from a given large set of handwritten

19
00:01:48,360 --> 00:01:51,660
digits, which is also known as the training examples.

20
00:01:53,250 --> 00:01:59,280
So basically we will take a large set of data feed into the neural network.

21
00:02:00,450 --> 00:02:08,310
Neural network will make some rules and we will use this brain neural network to predict for the other

22
00:02:08,310 --> 00:02:08,910
cases.

23
00:02:10,890 --> 00:02:17,890
This ability to create own rules makes neural network versatile, powerful and scalable.

24
00:02:19,590 --> 00:02:24,060
And this course, we are going to take the example of fashion m NASD.

25
00:02:27,770 --> 00:02:39,230
Fashion m NASD is a data set in which we classify images into 10 different fashion items in this dataset.

26
00:02:39,620 --> 00:02:48,380
We will have images of several fashion items and we will classify them into 10 different categories.

27
00:02:48,650 --> 00:02:52,460
For example, the first image is of her ankle boot.

28
00:02:53,720 --> 00:02:55,400
The second is a pullover.

29
00:02:56,180 --> 00:02:58,490
Third is a trouser and so on.

30
00:02:59,570 --> 00:03:07,310
So we will have grayscale images of several fashion items and will classify them into several fashion

31
00:03:07,340 --> 00:03:07,880
product.

32
00:03:11,090 --> 00:03:16,850
This problem statement is going to be significantly more challenging than the handwriting recognition

33
00:03:16,850 --> 00:03:17,360
example.

34
00:03:18,410 --> 00:03:22,940
And once you are done with this, you will appreciate the portal neural networks.

35
00:03:25,580 --> 00:03:30,290
Let me give you an overview of what the next few lectures are going to be.

36
00:03:32,120 --> 00:03:39,890
We will start the conceptual part by understanding two very important artificial neural cells called

37
00:03:40,160 --> 00:03:42,090
Perceptron and a sigmoid neuron.

38
00:03:43,790 --> 00:03:50,540
Then we will stack these single cells into multiple cells and we will make a multi-level Perceptron.

39
00:03:51,000 --> 00:03:55,280
A multi-level Perceptron is the most commonly used neural network model.

40
00:03:57,350 --> 00:04:01,460
Then we will dig deeper and understand forward and backward propagation.

41
00:04:02,600 --> 00:04:08,960
Then we move on to stochastic gradient descent, which is the standard learning algorithm for neural

42
00:04:08,960 --> 00:04:11,930
networks and protocols.

43
00:04:12,170 --> 00:04:15,560
We will be implementing all these models in Python.

44
00:04:17,810 --> 00:04:22,640
And this course, I will provide you with the reason of why we are doing whatever we are doing.

45
00:04:23,540 --> 00:04:30,770
So by the end of the course, you will understand what neural network and deep learning is, how to

46
00:04:30,770 --> 00:04:38,180
create a neural network to solve complex problems, and when and where to use neural networks.
