WEBVTT 00:02.130 --> 00:08.310 Hello and welcome to this tutorial here, we will understand, unless the young and dear, you let us 00:08.310 --> 00:08.970 begin. 00:10.580 --> 00:17.580 Alice D.M. stands for long short term memory, whereas GIU stands for gaited, Greek current. 00:17.740 --> 00:18.250 You need. 00:20.070 --> 00:22.510 I ran into a strong short term memory. 00:22.780 --> 00:30.580 If I think events is long enough, then adenine will have a hard time to get the information from earlier 00:30.580 --> 00:32.890 times stapes to the later ones. 00:33.820 --> 00:36.520 Let us understand this with the help of an example. 00:38.350 --> 00:45.490 If you are trying to process a paragraph of text to do predictions, then recurrent neural networks 00:45.490 --> 00:49.540 may leave out important information from the beginning. 00:52.800 --> 00:53.490 Let's deal. 00:53.670 --> 00:57.930 And Jerry, you were created as the solution to short term memory. 00:59.400 --> 01:02.020 They have internal mechanics called as Gates. 01:02.220 --> 01:04.800 They can regulate the flow of information. 01:07.970 --> 01:13.320 In this picture, we have two diagrams, unless the young and G are you left side? 01:13.400 --> 01:20.120 We help enlist the young and right side, we help guide you in LSP and there are four main components 01:20.290 --> 01:26.000 for Gate Gate, Sayle, gate output gate and input gate vair. 01:26.150 --> 01:27.320 Whereas in Jiahu. 01:27.350 --> 01:29.540 There are only two main components. 01:29.750 --> 01:31.790 It is said gate and update gate. 01:32.570 --> 01:37.040 At the bottom of these two diagrams you can see symbols and their meanings. 01:37.730 --> 01:43.100 First one is the sigmoid activation function, then tennet activation function. 01:43.310 --> 01:50.420 After that point, base multiplication, then point wise addition and at the end vector concatenation. 01:51.290 --> 01:57.470 The gate that you can see here in these two diagrams can learn which data elastic weights is important 01:57.470 --> 01:59.390 to keep and which is not. 01:59.900 --> 02:06.890 By doing that, these gate passes to relevant information down the long chain of thick vents to make 02:06.890 --> 02:07.940 deep predictions. 02:09.050 --> 02:11.220 Let us understand at least the first. 02:11.450 --> 02:18.820 As you can see here, in the end, there are four gates for gate output update and input. 02:19.250 --> 02:23.600 Forget Gate will decide what to forget from the previous memory you need. 02:24.230 --> 02:30.710 The input gate will decide what to accept inside the new Rhon update gate will update the memory IDs 02:31.100 --> 02:34.970 and the output gate will do the output as long term memory. 02:35.720 --> 02:37.980 So this is the working of Alist the young. 02:38.170 --> 02:38.980 Let us understand. 02:39.270 --> 02:46.820 Working off GIAT you Jiahu is a comparatively new generation of neural networks and it is similar to 02:46.880 --> 02:47.450 LSD. 02:48.440 --> 02:55.000 There are only two gates in GIAT you received and update gate the update gate access. 02:55.010 --> 03:00.020 Forget gate and input gate inside the LSP update gate. 03:00.050 --> 03:07.730 Besides what information to throw away and what information to add in, Meuron reset gate is used to 03:07.730 --> 03:10.620 decide how much pot information to for gate. 03:11.330 --> 03:18.350 So the city working off GIAT you using illest the young and yet you short term memory will be long term 03:18.350 --> 03:19.000 memory. 03:19.940 --> 03:23.760 So this ordeal about the Young and Jianyu NCR. 03:24.150 --> 03:26.990 I will see you in the next one till then. 03:27.150 --> 03:28.140 Happy learning.