1
00:00:11,100 --> 00:00:15,150
So in this lecture, we will be summarizing everything we learned in this section.

2
00:00:15,960 --> 00:00:18,720
This section was all about recurrent neural networks.

3
00:00:19,590 --> 00:00:23,040
As with CNN's, this section was really split into two parts.

4
00:00:23,460 --> 00:00:30,270
How Arnold's work and how to implement them in TensorFlow in terms of how organs work, we learned about

5
00:00:30,270 --> 00:00:35,310
three different variants the Alman unit, the GRU and the Alice team.

6
00:00:36,210 --> 00:00:43,140
In the past, the GRU was said to be on par with the LSHTM, but more recently the LSM has been shown

7
00:00:43,140 --> 00:00:44,760
to outperform the GRU.

8
00:00:46,520 --> 00:00:53,090
As such, are examples used the team by default, but you were encouraged to try other units as an exercise

9
00:00:53,900 --> 00:00:54,930
in terms of code.

10
00:00:54,950 --> 00:00:56,960
We looked at multiple examples.

11
00:00:57,590 --> 00:01:01,580
The first example was a simple example, which was to classify text.

12
00:01:02,090 --> 00:01:04,819
We did the same example with Anand's and CNN's.

13
00:01:05,570 --> 00:01:11,480
We learned that this is called a many to one task, since the input has many tokens, while the target

14
00:01:11,480 --> 00:01:12,500
only has one.

15
00:01:13,460 --> 00:01:17,390
We then looked at many too many tasks, such as parts of speech tagging.

16
00:01:18,170 --> 00:01:24,140
We observed that you could naively apply all the code we learned and get 99 percent accuracy.

17
00:01:24,860 --> 00:01:30,080
However, I also told you why this is misleading and showed you how to do this task correctly.

18
00:01:31,420 --> 00:01:38,350
Unfortunately, when you do write the code to do this correctly, TensorFlow is not as performance at

19
00:01:38,350 --> 00:01:38,990
this stage.

20
00:01:39,010 --> 00:01:45,070
Practically speaking, for the real world, it may be worth investing in PyTorch if you want to use

21
00:01:45,070 --> 00:01:47,130
these techniques on a real world project.

22
00:01:47,740 --> 00:01:52,750
Both have pros and cons, but if what you care about is speed, then that's always another option to

23
00:01:52,750 --> 00:01:53,380
explore.