1
00:00:11,070 --> 00:00:16,440
In this video, we are going to summarize everything we learned in the section and perhaps give you

2
00:00:16,440 --> 00:00:21,450
a few things to think about, this section focused on recurrent neural networks.

3
00:00:22,050 --> 00:00:27,450
As usual, we first learned about the theory behind Arnon's and then we use them in code four time series

4
00:00:27,450 --> 00:00:28,440
applications.

5
00:00:29,490 --> 00:00:34,830
One of the nice features of tens of flowing cameras is that they make working with Arnon's very, very

6
00:00:34,830 --> 00:00:35,460
easy.

7
00:00:36,090 --> 00:00:41,100
A lot of you new students won't realize this, but Arnon's are actually pretty complicated to build,

8
00:00:41,460 --> 00:00:43,740
much more so than Anand's or cornets.

9
00:00:44,490 --> 00:00:48,900
So if you ever took my original course on Arnon's, you've seen this for yourself.

10
00:00:49,470 --> 00:00:54,210
Basically, it's because of the fact that when you're working with a static computation graph, you

11
00:00:54,210 --> 00:00:58,090
cannot use for loops, at least not the kind you typically think of.

12
00:00:58,590 --> 00:01:03,740
And of course, for loops are kind of what you need in order to iterate over a sequence of inputs.

13
00:01:04,380 --> 00:01:08,930
In the old days, we use the YANNO, which was the original deep learning library in Python.

14
00:01:09,630 --> 00:01:14,470
And so if you ever built in Ahren and at V.A., then you've experienced what I'm referring to.

15
00:01:15,210 --> 00:01:20,130
Now, luckily, in modern times, you can still get that same experience if you like.

16
00:01:20,910 --> 00:01:25,080
It turns out that Google has not one but two deep learning libraries.

17
00:01:25,420 --> 00:01:31,170
One is obviously tensor flow, but the other is called Jacques's, which has been quite popular internally.

18
00:01:32,040 --> 00:01:36,840
So Jacques's is a lower level library, meaning that you typically implement things from scratch, the

19
00:01:36,840 --> 00:01:38,190
same as you wouldn't theno.

20
00:01:38,700 --> 00:01:44,460
And because of this very similar programming, constructs are used when you build in Arnet and Jaxx.

21
00:01:45,570 --> 00:01:50,460
So it's kind of an interesting aside on how you might get the full or an end experience.

22
00:01:55,080 --> 00:02:00,240
Now, what I hope you got out of this section is that it made you think think about what assumptions

23
00:02:00,240 --> 00:02:02,350
people make when it comes to our own ends.

24
00:02:02,730 --> 00:02:08,490
I see a lot of students get really excited about Arnon's and Steam's, but keep in mind that Time series

25
00:02:08,640 --> 00:02:10,740
are not where the LSM proved itself.

26
00:02:11,700 --> 00:02:17,500
In fact, it appears to me that some students get away more excited about ellis' times than they should.

27
00:02:17,910 --> 00:02:21,710
I've seen many instances where CNN's outperform Ms.

28
00:02:22,440 --> 00:02:23,820
So what's the lesson here?

29
00:02:24,990 --> 00:02:29,270
Well, the lesson is that don't use something simply because you think it sounds cool.

30
00:02:29,520 --> 00:02:31,470
Instead, use what works.

31
00:02:32,700 --> 00:02:37,560
Of course, the only way to find out whether or not something will work is to run experiments and to

32
00:02:37,560 --> 00:02:38,810
try those things yourself.

33
00:02:39,420 --> 00:02:43,500
That said, what we did in this section does count as running experiments.

34
00:02:43,710 --> 00:02:49,110
So add what you've observed here to your list of experience and use that to guide your future decisions

35
00:02:49,110 --> 00:02:52,220
about whether or not you will choose to use the MS.