1
00:00:00,930 --> 00:00:01,320
Hello.

2
00:00:01,350 --> 00:00:02,490
Welcome back.

3
00:00:02,490 --> 00:00:08,810
In this lesson we are going to talk about the learn in part of machine learning and machine learning

4
00:00:08,820 --> 00:00:09,960
when we say learning.

5
00:00:10,020 --> 00:00:13,040
We simply mean reducing error.

6
00:00:13,710 --> 00:00:21,240
Learning involves calculating the direction and amount by which we need to change the weight in order

7
00:00:21,240 --> 00:00:26,350
to reduce that difference between the predicted value and the expected value.

8
00:00:26,400 --> 00:00:33,160
Ideally our goal is to make the predicted value as close to the expected value as possible.

9
00:00:33,270 --> 00:00:39,000
As we mentioned earlier the difference between the predicted value and the expected value is what we

10
00:00:39,000 --> 00:00:40,430
call error.

11
00:00:40,470 --> 00:00:44,670
We square this difference to ensure that it is always positive.

12
00:00:44,670 --> 00:00:49,350
We are going to introduce a new variable known as Arrow direction and amount.

13
00:00:49,350 --> 00:00:56,580
I know that's a long name for a variable but it is very descriptive to find the error direction and

14
00:00:56,640 --> 00:01:04,650
amount we subtract the expected value from the predicted value and then we multiply the answer by the

15
00:01:04,710 --> 00:01:06,600
inputs.

16
00:01:06,600 --> 00:01:13,630
We then update the weight by so cracked in the error direction and amount from the weight.

17
00:01:13,680 --> 00:01:21,120
The answer is the new weight this new weight is what is multiplied by the inputs to get the next predicted

18
00:01:21,120 --> 00:01:28,960
value predicted value minus expected value is what is known as Pure error.

19
00:01:29,010 --> 00:01:32,880
Note that over here we do not square the answer.

20
00:01:32,890 --> 00:01:38,120
The pilot error tells us the direction and how much we missed by direction.

21
00:01:38,130 --> 00:01:46,710
I mean whether positive or negative if the pilot error is a positive number it means we predicted too

22
00:01:46,710 --> 00:01:47,850
high.

23
00:01:47,880 --> 00:01:53,950
If the pure error is a negative number it means we predicted too low.

24
00:01:54,000 --> 00:01:59,870
If it is a big number it means we missed by a big amount.

25
00:01:59,940 --> 00:02:05,890
If it is a big number I repeat it means we missed by a large or big amount.

26
00:02:05,970 --> 00:02:11,980
We multiply by the input to get three effect.

27
00:02:12,000 --> 00:02:14,970
If the input is zero then it will force the error.

28
00:02:15,000 --> 00:02:20,890
Difference in amount variable to be zero because there is nothing to learn.

29
00:02:21,900 --> 00:02:28,680
Multiplying the pure arrow by the input will reverse the sine and direction of the arrow direction an

30
00:02:28,680 --> 00:02:34,620
amount variable in the situation where the M word input is negative.

31
00:02:34,800 --> 00:02:44,820
If the input is a pig we want the weight update to be big as well and we get this scale in effect also

32
00:02:44,820 --> 00:02:48,040
by multiplying the pure arrow by the input.

33
00:02:48,300 --> 00:02:55,700
Because this third effect could get out of control we will introduce a new constants known as the land

34
00:02:55,770 --> 00:03:03,350
rate which is denoted by Alpha to prevent the scale in effect.

35
00:03:03,360 --> 00:03:07,440
Now let's update our pseudocode to introduce alpha

36
00:03:12,530 --> 00:03:18,500
in this sort of code rather than compute the error amount and direction in one line.

37
00:03:18,560 --> 00:03:24,710
We have broken it into two lines and also introduce two new verbose known as Delta.

38
00:03:24,830 --> 00:03:26,770
And wait Delta.

39
00:03:27,380 --> 00:03:33,380
We find Delta by subtracting the expected value from the predicted value.

40
00:03:33,530 --> 00:03:41,150
Delta is computed by multiplying the data by the input in effect weights.

41
00:03:41,150 --> 00:03:48,170
Delta is the same as what we previously had as error direction and amount.

42
00:03:48,170 --> 00:03:56,330
Now we update the weight by multiplying debate Delta by D constant Alpha which is the land and rate

43
00:03:56,900 --> 00:04:00,760
and then subtract in the answer from the current weight.

44
00:04:01,010 --> 00:04:05,230
The answer becomes the updated weight over here.

45
00:04:05,240 --> 00:04:14,140
We simply use in alpha equals zero point zero one in which I'll experiment with these values programmatically.

46
00:04:14,960 --> 00:04:16,520
So that's all there is for this lesson.

47
00:04:16,520 --> 00:04:19,490
If you have any questions at all just let me know.

48
00:04:19,550 --> 00:04:20,240
I'll see you later.