1
00:00:02,260 --> 00:00:05,800
Lane it is a very simple architecture.

2
00:00:05,910 --> 00:00:13,260
It has three convoluted layers two of the Congressional leaders also have average pooling layer.

3
00:00:13,510 --> 00:00:20,370
Notice that it is not Max pooling it was average pulling as I told you earlier also average pooling

4
00:00:20,370 --> 00:00:25,500
was more popular in the earlier days and later on Max pooling became more popular.

5
00:00:26,280 --> 00:00:32,670
So limit being one of the ugliest congressional neural network architectures had average pooling.

6
00:00:34,020 --> 00:00:41,840
So the first convolution led and training conversion led have average pooling third conversion led straight

7
00:00:41,840 --> 00:00:47,220
up it gives its output to a fully connected neural network.

8
00:00:47,280 --> 00:00:54,370
If you look at the input image it took in an image of size 32 by 32.

9
00:00:55,170 --> 00:01:05,430
And after the convolution layers we had 120 such features of one by one size these would fit into a

10
00:01:05,430 --> 00:01:09,150
fully connected neural network.

11
00:01:09,150 --> 00:01:17,370
This problem was run on amnesty only which is handwriting recognition data and it was able to achieve

12
00:01:17,370 --> 00:01:19,970
very good accuracies.

13
00:01:20,040 --> 00:01:21,840
This architecture is very simple.

14
00:01:21,870 --> 00:01:27,750
In fact I would encourage you to make this architecture in your system you know everything that you

15
00:01:27,750 --> 00:01:30,720
need to know to create this architecture and you can run this.