1
00:00:00,570 --> 00:00:00,980
Hello.

2
00:00:01,020 --> 00:00:01,890
Welcome back.

3
00:00:01,890 --> 00:00:05,770
Let's see what one layer of convolution looks like.

4
00:00:05,850 --> 00:00:14,650
So our input is denoted over here by a superscript 0 because the input layer is often called Layer 0.

5
00:00:14,880 --> 00:00:18,650
Our weight is denoted by top use superscript 1.

6
00:00:19,200 --> 00:00:25,530
So to compute for 0 1 we can form the input with the weight and then we add the buyers.

7
00:00:25,530 --> 00:00:29,220
B this gives us c 1.

8
00:00:29,220 --> 00:00:38,010
So z 1 equals w 1 which is the weight convolution a 0 which is the input plus B which is the bias to

9
00:00:38,010 --> 00:00:46,310
compute the activation a one we simply take c 1 and apply a 1 activation function.

10
00:00:46,320 --> 00:00:50,320
Now let's see how to find the number of parameters of a layer.

11
00:00:50,340 --> 00:00:57,220
Let's say we have 10 failed tests and each filter has a shape of three by three by three.

12
00:00:57,570 --> 00:01:05,600
We find a number of parameters per filter by simply doing three multiplied by three which apply by three.

13
00:01:06,990 --> 00:01:12,560
And this of course this would give us twenty seven.

14
00:01:12,600 --> 00:01:19,620
Also there would be a bias term for each field to be the bias each filter has a bias.

15
00:01:19,620 --> 00:01:21,750
So we have to add one to it.

16
00:01:22,020 --> 00:01:26,590
So twenty seven plus one gives us twenty eight.

17
00:01:26,610 --> 00:01:33,960
And because we are below with 10 filters then we have to multiply what twenty eight by ten and this

18
00:01:33,960 --> 00:01:36,440
gives us two hundred and eighty.

19
00:01:36,690 --> 00:01:41,670
And as you can see from this calculation the only thing taken into account when finding the number of

20
00:01:41,670 --> 00:01:45,270
parameters is the shape and number of filters.

21
00:01:45,330 --> 00:01:53,180
This means no matter how big the input image is the parameters only depend on the filter size.

22
00:01:53,190 --> 00:01:59,580
Let's take a look at the notation we shall use or the notations which will use when talking about convolution

23
00:01:59,580 --> 00:02:08,220
or new network layer is denoted as L such that when we want to talk about the filter size that pattern

24
00:02:08,220 --> 00:02:15,870
described the number of channels etc. we simply write the respective letters and add superscript square

25
00:02:15,900 --> 00:02:23,220
brackets l the previous layer is the input of the current layer therefore we use L minus 1 to denote

26
00:02:23,310 --> 00:02:29,580
input H and W denote height and with respectively.

27
00:02:29,580 --> 00:02:34,120
This is all there is for this lesson and if you have any questions just let me know.

28
00:02:34,170 --> 00:02:35,210
I shall see you later.