1
00:00:02,860 --> 00:00:06,390
Now let us look at another CNN architecture.

2
00:00:06,460 --> 00:00:14,200
Google it was the winner of 2014 challenge Google and had one new concept and the concept was often

3
00:00:14,350 --> 00:00:25,600
Inception module Inception Module looks something like this the input to the inception module is given

4
00:00:26,050 --> 00:00:33,520
to four different layers three of these layers are convolution layers and the fourth one is a max puller.

5
00:00:35,740 --> 00:00:42,640
If you look at these convolution layers these are also one by one kernel that is the window is of the

6
00:00:42,640 --> 00:00:44,860
size of a single pixel.

7
00:00:45,100 --> 00:00:50,230
Usually we have been using continuously layers with two by two or three by 2.0.

8
00:00:50,980 --> 00:00:56,140
But in the inception module you can see that Windows or is one by one.

9
00:00:56,230 --> 00:01:03,880
Used this oneness represents this trade so it has a straight of one the output of these two convolution

10
00:01:03,900 --> 00:01:09,910
alert and this match pulling layer then went into three different congressional layers.

11
00:01:10,090 --> 00:01:18,210
The output of these four was then put into a debt concatenated will not be discussing that concatenated

12
00:01:18,220 --> 00:01:19,850
here.

13
00:01:19,870 --> 00:01:28,510
This whole thing is called an inception module and the actual architecture of Google it was something

14
00:01:28,510 --> 00:01:32,830
like this input of images went through here.

15
00:01:32,830 --> 00:01:39,640
Then there was a convolution layer a max puller a local response normally to convolution layers and

16
00:01:39,640 --> 00:01:40,700
so on.

17
00:01:40,870 --> 00:01:44,220
All of these are staggered Inception modules.

18
00:01:44,320 --> 00:01:46,120
So this is the inception layer.

19
00:01:46,120 --> 00:01:50,680
This is an separate layer so many of these Inception layers.

20
00:01:51,070 --> 00:01:57,550
The output of these goes on to another set of Inception layers and then we finally have a fully connected

21
00:01:58,060 --> 00:01:58,840
neural network

22
00:02:01,740 --> 00:02:10,800
so if we look at it it is a very complex network although it had very few training parameters for mostly

23
00:02:10,860 --> 00:02:16,590
working professionals or students working in the field of data science and machine learning.

24
00:02:16,590 --> 00:02:22,650
Creating such architectures on their machines is not possible.

25
00:02:22,650 --> 00:02:32,030
So one of the good things that is coming with our gate US library is that we are able to use these pre

26
00:02:32,040 --> 00:02:35,460
trained models for our problem

27
00:02:38,730 --> 00:02:47,250
these architectures were created to solve one particular problem but tenants allow us to use these architectures

28
00:02:48,240 --> 00:02:49,380
for other problems.

29
00:02:49,380 --> 00:02:51,660
Also let us see how.