1
00:00:02,780 --> 00:00:05,960
Next is Dee Dee Dee 16.

2
00:00:06,020 --> 00:00:11,590
This was the runner up in 2014 a year less VRC challenge.

3
00:00:11,720 --> 00:00:19,670
This was a very big architecture and it had 138 million parameters to be trained so it took a lot of

4
00:00:19,670 --> 00:00:21,920
time to train.

5
00:00:22,010 --> 00:00:29,570
If you look at the architecture instead of convoluted many layers it has conventional blocks and one

6
00:00:29,570 --> 00:00:37,640
block consists of several conventional layers and a max pooling layer at the end to this constitutional

7
00:00:37,640 --> 00:00:47,000
block has to convolution in layers and one Max pooling layer next conditional block has two corners

8
00:00:47,000 --> 00:00:53,270
layers and a max willingly and then we again have a congressional block.

9
00:00:53,720 --> 00:00:59,630
The last two convoluted blocks have three constitutional layers instead of two congressional leaders

10
00:01:02,090 --> 00:01:07,060
after processing up to 20 forward to 24 pixel image.

11
00:01:07,520 --> 00:01:19,050
At the end of all of these convolution layers we had 512 feature maps of 7 by 7 size these feature maps

12
00:01:19,140 --> 00:01:28,810
were fed into a fully connected neural network with so many neurons at the end of it.

13
00:01:28,950 --> 00:01:35,550
There were 1000 output neurons with soft maps activation.

14
00:01:35,580 --> 00:01:43,830
This is because we did see sixteen was run on a problem in which there were 1000 glasses into which

15
00:01:43,950 --> 00:01:48,220
images were to be classified so it was a huge database.

16
00:01:48,480 --> 00:01:56,370
Over a million images and those images were to be classified into one thousand Glasses glasses were

17
00:01:56,430 --> 00:01:59,250
of dog breeds God breeds etc..