1
00:00:00,528 --> 00:00:04,610
Last week, we looked at recurrent neural
networks including a simple RNN and

2
00:00:04,610 --> 00:00:05,250
an LSTM.

3
00:00:06,330 --> 00:00:09,530
You saw how they could be useful in
learning a sequence like the one that

4
00:00:09,530 --> 00:00:11,020
we've been looking at.

5
00:00:11,020 --> 00:00:14,590
And how the LSTMs removed some of
the issues we had with an RNN.

6
00:00:15,750 --> 00:00:17,590
This week, you'll go one step further,

7
00:00:17,590 --> 00:00:22,054
combining convolutions with LSTMs
to get a very nicely fitting model.

8
00:00:22,054 --> 00:00:25,630
We'll then apply that to real
world data instead of this

9
00:00:25,630 --> 00:00:28,970
synthetic data set that we've been working
on since the beginning of this course.

10
00:00:31,620 --> 00:00:33,990
Here's the LSTM that you
looked at last week,

11
00:00:33,990 --> 00:00:37,320
except that I've added something at
the beginning of the sequential stack.

12
00:00:38,570 --> 00:00:42,570
It's a convo D where we'll
try to learn 32 filters.

13
00:00:42,570 --> 00:00:44,070
It's a one dimensional convolution.

14
00:00:44,070 --> 00:00:48,070
So we'll take a five number window and
multiply out the values in that

15
00:00:48,070 --> 00:00:52,630
window by the filter values, in much the
same way as image convolutions are done.

16
00:00:54,500 --> 00:00:56,850
Check back to the convolutions
course to learn more about that.