1 00:00:11,130 --> 00:00:17,100 So in this video, we will summarize everything we learned in this section, this section introduced 2 00:00:17,280 --> 00:00:19,020 artificial neural networks. 3 00:00:19,860 --> 00:00:25,110 We learned how to build a basic feed for it and then using tensor flow, but more importantly, how 4 00:00:25,110 --> 00:00:27,430 to build more complex architectures as well. 5 00:00:28,080 --> 00:00:33,780 Specifically, we learned how to build a multi input neural network to handle multivariate time series 6 00:00:33,900 --> 00:00:35,220 in a naive way. 7 00:00:36,120 --> 00:00:41,250 We also learned some of the important theory behind neural networks, such as the neuron analogy and 8 00:00:41,250 --> 00:00:45,000 the power of adding more layers with nonlinear activations. 9 00:00:46,020 --> 00:00:48,120 We then discovered what it means to learn. 10 00:00:48,660 --> 00:00:52,880 Although this sounds somewhat magical, it is in fact just a calculus problem. 11 00:00:53,370 --> 00:00:56,610 More specifically, learning is just optimization. 12 00:01:01,270 --> 00:01:06,970 In this section, we apply to deep learning to multiple data sets, we, of course, started with airline 13 00:01:06,970 --> 00:01:09,930 passengers, which has been our benchmark for this cause. 14 00:01:10,420 --> 00:01:15,550 We also apply to deep learning to stock returns, although, as usual, the results may have been a 15 00:01:15,550 --> 00:01:16,730 hard truth to some. 16 00:01:17,890 --> 00:01:22,350 We then learned about a new data set, which is especially well suited to deep learning. 17 00:01:23,140 --> 00:01:25,320 This was human activity recognition. 18 00:01:25,840 --> 00:01:31,230 This is a multiclass classification problem with the multivariate time series as input. 19 00:01:32,590 --> 00:01:38,680 We saw that the multi tailed A&M is useful for processing multivariate time series and can handle them 20 00:01:38,680 --> 00:01:39,930 in a non-native way. 21 00:01:40,660 --> 00:01:45,910 The naive way would be to simply flatten all the data into a single feature vector, but that wouldn't 22 00:01:45,910 --> 00:01:48,820 take advantage of the structure in the Time series. 23 00:01:50,020 --> 00:01:55,210 Now, despite the promise of deep learning, being able to automatically find features, we learn that 24 00:01:55,210 --> 00:01:58,660 sometimes creating features ourselves can still be useful. 25 00:01:59,470 --> 00:02:05,050 We saw that by using predefined features of our Time series data, we were able to obtain a superior 26 00:02:05,050 --> 00:02:05,900 performance. 27 00:02:06,850 --> 00:02:12,760 We also learned how to combine static features with the Time series model to obtain even better performance.