1 00:00:11,140 --> 00:00:16,420 OK, so we are now entering the next section of this course, so this section is where we begin our 2 00:00:16,420 --> 00:00:19,370 study of deep learning for a time series analysis. 3 00:00:20,170 --> 00:00:24,280 We'll start with learning about Anan's, which stands for artificial neural networks. 4 00:00:24,910 --> 00:00:30,330 Now, technically, if all you wanted to do was use a simple and then you wouldn't really need the section, 5 00:00:30,730 --> 00:00:32,230 you could just use the MLP. 6 00:00:32,230 --> 00:00:33,280 And I learn. 7 00:00:34,120 --> 00:00:39,610 But with deep learning, it's worth knowing about how it actually works so that you can go beyond simple 8 00:00:39,610 --> 00:00:42,970 and ends up to more complex architectures such as CNN. 9 00:00:42,970 --> 00:00:46,220 Zanón ends for this will be using Tenzer flow. 10 00:00:46,900 --> 00:00:50,940 So essentially the deep learning sections boil down to two components. 11 00:00:51,640 --> 00:00:52,030 No. 12 00:00:52,030 --> 00:00:54,790 One, understanding how and why the model works. 13 00:00:55,000 --> 00:00:58,570 And number two, how to implement it using tensor flow. 14 00:01:03,320 --> 00:01:05,690 So the outline for this section is as follows. 15 00:01:06,260 --> 00:01:09,530 We'll start with understanding the basic concept of the neuron. 16 00:01:10,310 --> 00:01:15,470 Basically, this isn't anything new in terms of models, but it's a new perspective on how to look at 17 00:01:15,470 --> 00:01:17,720 the linear models we've already used. 18 00:01:18,710 --> 00:01:21,920 The next step is to combine neurons, to get neural networks. 19 00:01:22,490 --> 00:01:28,280 Once we understand that, we'll learn about how to think of what animals do from a geometrical perspective. 20 00:01:28,940 --> 00:01:35,270 That is why our neural networks, nonlinear, will then discuss different activation functions and how 21 00:01:35,270 --> 00:01:38,210 neural networks can be used for multiclass problems. 22 00:01:39,590 --> 00:01:43,050 The next step will be to use interns for multiple data sets. 23 00:01:43,610 --> 00:01:47,870 We'll start with airline passengers, which has been our benchmark data set for this course. 24 00:01:48,440 --> 00:01:52,850 We'll also look at stock predictions and we'll see that this doesn't require any new code. 25 00:01:54,050 --> 00:02:00,450 The next step will be to look at a new data set called Human Activity Recognition for this data set. 26 00:02:00,650 --> 00:02:06,110 We'll get to build a time series classifier, which can take as input a multivariate time series. 27 00:02:06,620 --> 00:02:08,660 This is significant for two reasons. 28 00:02:09,470 --> 00:02:15,710 Firstly, we haven't yet seen any classifiers that can handle multivariate time series in a non way. 29 00:02:16,370 --> 00:02:21,170 Theoretically, you could just use the machine learning models we discussed previously, but that would 30 00:02:21,170 --> 00:02:26,820 be naive since it treats every sample from every timestep and every channel the same way. 31 00:02:27,890 --> 00:02:32,720 This leads us to our second point, which is that in order to solve this problem, you'll need to learn 32 00:02:32,720 --> 00:02:34,820 about a new neural network architecture. 33 00:02:35,450 --> 00:02:38,570 OK, so this is not your typical intro to Andsnes. 34 00:02:38,900 --> 00:02:44,600 This will include a new way to build in ends that can handle multivariate time series, not naively. 35 00:02:49,200 --> 00:02:54,000 As a final note for this lecture, I want to mention that my sense of the two course is what I call 36 00:02:54,000 --> 00:02:56,840 a soft prerequisite for the deep learning sections. 37 00:02:57,240 --> 00:03:01,050 It's not critical because we will go over all the necessary details. 38 00:03:01,440 --> 00:03:05,880 However, if you want to have a better understanding of what we are doing, then you'll want to have 39 00:03:05,880 --> 00:03:06,930 that experience. 40 00:03:08,250 --> 00:03:14,160 Alternatively, my torch chorus or my in-depth deep learning series are also viable prerequisites. 41 00:03:14,760 --> 00:03:19,320 And this is because if you've already learned PI talk, you should find tensor flow to quite easy.