WEBVTT

00:02.150 --> 00:06.520
Hello and welcome to Digital in Digital Delivery, understand?

00:07.100 --> 00:09.010
Vanishing gradient problem.

00:09.260 --> 00:10.220
Let us begin.

00:11.870 --> 00:18.890
When we do back propagation that is moving backward in network and calculating gradient of laws with

00:18.890 --> 00:26.000
respect to Veidt, then the gradient tends to get smaller and smaller as we keep on moving backward

00:26.060 --> 00:26.980
in the network.

00:28.580 --> 00:35.090
That means the neutrons in earlier layers launched very slowly as compared to the neutrons that are

00:35.150 --> 00:38.990
in literally us, the earlier layers in the network.

00:39.190 --> 00:40.670
This law is to train.

00:41.420 --> 00:48.740
So the process of reducing gradient as we go backward in network is called as vanishing gradient problem.

00:51.820 --> 00:59.220
Let us understand the importance of earlier, least in neural network, early, early hours in the network

00:59.390 --> 01:05.960
are important because they are responsible to learn and detect simple pat downs and they are actually

01:05.960 --> 01:08.180
the building blocks of our network.

01:09.770 --> 01:17.150
If these earlier layers will do improper and inaccurate results, then how can we expect that the next

01:17.150 --> 01:20.870
layers to perform nicely and produce accurate results?

01:21.410 --> 01:25.390
So this is the importance of earlier layers in neural network.

01:28.530 --> 01:31.950
Let us discuss dissolution to varnishing gradient problem.

01:33.680 --> 01:41.180
We do not use sigmoid Antonette as an activation function which caused this vanishing gradient problems,

01:42.890 --> 01:50.060
instead of that, we use RELU based activation functions in training deep neural network model to avoid

01:50.060 --> 01:51.320
such complications.

01:51.560 --> 01:53.570
And we can improve the accuracy.

01:54.500 --> 01:59.630
So theoretically, this is all about the vanishing gradient problem and its solution.

01:59.930 --> 02:02.830
I will see you in the next door deal till then.

02:03.170 --> 02:04.190
Happy learning.