1 00:00:02,100 --> 00:00:07,180 Once we've studied Arnon's, we can go even deeper into how deep learning can be used on text. 2 00:00:07,590 --> 00:00:13,320 That's the next course, natural language processing with deep learning in Python or deep learning in 3 00:00:13,320 --> 00:00:14,520 Python, part six. 4 00:00:15,440 --> 00:00:20,840 In this course, we look at a very important concept called a Werdum betting's, these allow us to turn 5 00:00:20,840 --> 00:00:26,400 words which are categorical variables into vectors, which are numbers that in turn that word can read. 6 00:00:27,260 --> 00:00:32,000 This is why it's very important to have studied unsupervised learning before this, because finding 7 00:00:32,000 --> 00:00:34,670 Werdum betting's is actually an unsupervised task. 8 00:00:35,360 --> 00:00:40,400 Word embedding also allow us to make use of pre training, which was discussed in deep learning in Python 9 00:00:40,400 --> 00:00:41,030 part for. 10 00:00:42,020 --> 00:00:47,510 In this course, we also look at a very advanced model for doing sentiment analysis called a recursive 11 00:00:47,510 --> 00:00:49,430 neural network or a tree neural network. 12 00:00:50,180 --> 00:00:55,580 This is an example of a dynamic neural network because it changes its structure based on what input 13 00:00:55,580 --> 00:00:56,090 you give it. 14 00:00:56,900 --> 00:01:00,920 Most deep learning libraries are not equipped to handle dynamic neural networks. 15 00:01:00,930 --> 00:01:06,740 And I demonstrate what happens if you try to build one the naive way you basically end up building a 16 00:01:06,740 --> 00:01:11,420 separate neural network for each of your training samples and it's going to eat up all your ram and 17 00:01:11,420 --> 00:01:13,350 make your computer slow down to a crawl. 18 00:01:14,150 --> 00:01:19,280 So you get a firsthand perspective of why working with dynamic neural networks is not easy. 19 00:01:20,000 --> 00:01:24,430 But luckily we are able to make use of our knowledge of recurrent neural networks. 20 00:01:24,440 --> 00:01:30,860 And remember, that's the prerequisite to this course in order to convert a tree into a sequence which 21 00:01:30,860 --> 00:01:32,820 Anadan is capable of handling. 22 00:01:33,350 --> 00:01:36,350 So that's why Arnon's is a prerequisite to this course. 23 00:01:38,250 --> 00:01:42,480 Now, you notice that this course, deep in L.P, has no outgoing edges. 24 00:01:42,510 --> 00:01:45,830 This is because it's the most advanced course that I have on this path. 25 00:01:45,840 --> 00:01:49,780 For now, I certainly plan on expanding in this direction in the future. 26 00:01:50,460 --> 00:01:54,180 However, this is not the last deep learning course in the series. 27 00:01:54,810 --> 00:01:59,910 After this, we have deep reinforcement learning, which is deep learning in Python part seven. 28 00:02:00,660 --> 00:02:05,430 With that said, I think now is a good time to go back and start exploring the reinforcement learning 29 00:02:05,430 --> 00:02:05,780 path. 30 00:02:07,020 --> 00:02:11,220 So you'll notice that I have Bayesian machine learning, feeding into reinforcement learning. 31 00:02:12,550 --> 00:02:17,740 On the surface, these two corpses might seem unrelated, but there is a very important concept you'll 32 00:02:17,740 --> 00:02:23,350 learn that applies to both called the explore exploit dilemma in Bayesian machine learning. 33 00:02:23,380 --> 00:02:28,480 You learn this idea in the context of trying to optimize it, click through rate or a conversion rate, 34 00:02:28,960 --> 00:02:33,850 or in other words, the number of times people buy things from your website versus the number of times 35 00:02:33,850 --> 00:02:35,180 people visited your website. 36 00:02:35,650 --> 00:02:36,880 Very practical concept. 37 00:02:36,880 --> 00:02:43,000 I think if you do anything related to e-commerce now in reinforcement learning, we look at the explore 38 00:02:43,000 --> 00:02:44,350 exploit dilemma again. 39 00:02:44,380 --> 00:02:50,020 But in the context of playing games, reinforcement learning is like a third branch of machine learning, 40 00:02:50,020 --> 00:02:53,600 whereas the other two are supervised and unsupervised learning. 41 00:02:54,280 --> 00:03:00,010 The main difference is that supervised and unsupervised learning look at static data and reinforcement 42 00:03:00,010 --> 00:03:00,360 learning. 43 00:03:00,370 --> 00:03:04,300 The idea is more like you have a robot living in the real world. 44 00:03:04,510 --> 00:03:10,350 It can take the experiences it had today and based on them, behave more intelligently tomorrow. 45 00:03:10,840 --> 00:03:14,030 So the learning paradigm and reinforcement learning is sequential. 46 00:03:14,650 --> 00:03:20,290 This is supposed to supervise an unsupervised learning where your data set usually resides in some file 47 00:03:20,290 --> 00:03:21,220 on your hard drive. 48 00:03:22,370 --> 00:03:28,370 So in reinforcement learning, part one, we get all the basics, as you might expect, this is a prerequisite 49 00:03:28,370 --> 00:03:33,890 to deep reinforcement learning, since deep reinforcement learning applies those concepts to more difficult 50 00:03:33,890 --> 00:03:34,490 games. 51 00:03:35,220 --> 00:03:39,290 But you'll notice that this is not the only prerequisite to deep reinforcement learning. 52 00:03:40,100 --> 00:03:41,440 As the title suggests. 53 00:03:41,450 --> 00:03:47,720 This is also dependent on knowledge of deep learning, in particular convolutional neural networks. 54 00:03:49,550 --> 00:03:54,920 But as we know, in order to build a CNN, we have to know how to build a regular neural network, which 55 00:03:54,920 --> 00:03:58,350 means we have to know what a neural network is and why it's useful and so on. 56 00:03:58,940 --> 00:04:02,330 So in deep reinforcement learning, these two paths converge. 57 00:04:02,750 --> 00:04:07,610 You combine your knowledge of both reinforcement learning and deep learning in this course. 58 00:04:08,890 --> 00:04:16,450 Now, the reason it depends on CNN's and not, say, Arnon's, which is over here, is because we'll 59 00:04:16,450 --> 00:04:18,250 be learning to play visual games. 60 00:04:18,250 --> 00:04:23,800 So, for example, we can learn how to play a video game like Pong, a breakout, which are classic 61 00:04:23,800 --> 00:04:25,320 Atari games from the old days. 62 00:04:25,690 --> 00:04:29,830 And so those are images because they're basically screenshots from the screen. 63 00:04:30,760 --> 00:04:36,550 In the future, we might end up applying Arnon's, in which case Arnon's will become a prerequisite 64 00:04:36,550 --> 00:04:37,300 to that course. 65 00:04:39,050 --> 00:04:41,900 So that's the end of the reinforcement learning path for now. 66 00:04:42,110 --> 00:04:45,880 I'm very excited to bring you more updates in this area in the future. 67 00:04:48,280 --> 00:04:50,890 Let's now jump back to you, logistic regression. 68 00:04:52,000 --> 00:04:58,600 Where we can see another outgoing edge to supervise machine learning, so why is this edge here? 69 00:04:59,490 --> 00:05:04,620 Well, you may recall that linear regression and logistic regression are both linear models that do 70 00:05:04,620 --> 00:05:06,870 regression and classification, respectively. 71 00:05:07,560 --> 00:05:09,690 These are both supervised learning tasks. 72 00:05:11,700 --> 00:05:17,130 And so it makes sense that now that, you know, one model for progression and one model for classification, 73 00:05:17,520 --> 00:05:20,080 it's time to dig deeper into supervised learning. 74 00:05:20,940 --> 00:05:25,730 The thing with linear regression and logistic regression is that they aren't really different models. 75 00:05:25,740 --> 00:05:29,130 They're both just the line because they do different tasks. 76 00:05:29,130 --> 00:05:32,040 The techniques and interpretations are slightly different to. 77 00:05:33,120 --> 00:05:37,410 And there are, of course, different models that are not linear models that can do these tests, and 78 00:05:37,410 --> 00:05:39,390 that's what this course is all about. 79 00:05:40,140 --> 00:05:45,750 In this course, we look at classic supervised machine learning techniques like Canibus Neighbor Decision 80 00:05:45,750 --> 00:05:48,570 Trees, the Perceptron and the baize classifier. 81 00:05:49,530 --> 00:05:54,990 Much like how logistic regression was the basic building block of the neural network, classic models 82 00:05:54,990 --> 00:05:58,800 like decision trees are the basic building block of ensemble methods. 83 00:05:59,780 --> 00:06:06,710 So that's why this course is a prerequisite to ensemble machine learning, again, we use the same logo 84 00:06:06,710 --> 00:06:11,480 with a different color to signify that these two courses are very closely related. 85 00:06:12,440 --> 00:06:14,010 In ensemble machine learning. 86 00:06:14,030 --> 00:06:19,640 We learn how to combine multiple decision trees in different ways in order to make some very powerful 87 00:06:19,640 --> 00:06:20,530 classifiers. 88 00:06:21,260 --> 00:06:26,190 What's really remarkable about these methods is that they are very easy to plug and play on data. 89 00:06:26,660 --> 00:06:31,820 So if you're looking for a plug and play solution without having to learn a lot of theory, then deep 90 00:06:31,820 --> 00:06:33,830 learning is most likely not for you. 91 00:06:34,190 --> 00:06:36,140 But ensemble methods are a great fit. 92 00:06:37,090 --> 00:06:41,620 Deep learning is very dependent on hyper parameters, and if you choose incorrectly, your model will 93 00:06:41,620 --> 00:06:42,790 perform very poorly. 94 00:06:43,420 --> 00:06:47,360 Sometimes it requires immense computing power to find good hyper parameters. 95 00:06:47,890 --> 00:06:50,860 This is an active area of research that has not yet solved. 96 00:06:51,670 --> 00:06:55,210 This is why you can implement what you see in a deep learning paper. 97 00:06:55,240 --> 00:07:00,760 But suppose the author left out some seemingly insignificant detail so you end up having to make an 98 00:07:00,760 --> 00:07:03,400 assumption and then your results end up totally different. 99 00:07:03,850 --> 00:07:08,060 So deep learning is fragile, but luckily ensemble methods are not. 100 00:07:08,770 --> 00:07:13,560 We focus on two very famous ensemble methods, the random forests and the boost. 101 00:07:14,380 --> 00:07:17,180 So that's everything on the supervised machine learning track. 102 00:07:17,200 --> 00:07:17,800 For now. 103 00:07:19,060 --> 00:07:25,720 Next, we see an edge going from supervised machine learning to unsupervised machine learning, in particular 104 00:07:25,720 --> 00:07:26,870 cluster analysis. 105 00:07:27,730 --> 00:07:32,920 The reason we study supervised learning before unsupervised learning is because unsupervised learning 106 00:07:32,920 --> 00:07:34,360 is a little more abstract. 107 00:07:34,900 --> 00:07:40,300 It takes more effort on the student's part to realize why it's practical and what it can be used for. 108 00:07:41,290 --> 00:07:47,210 Cluster analysis shows us how to model data that does not come with targets, as you might guess. 109 00:07:47,230 --> 00:07:49,150 We do this in the form of clustering. 110 00:07:49,930 --> 00:07:52,060 The idea behind clustering is very simple. 111 00:07:52,630 --> 00:07:57,580 We want to know how many naturally occurring groups of data there are and what are the relationships 112 00:07:57,580 --> 00:07:59,390 between the data in these clusters. 113 00:07:59,800 --> 00:08:04,900 So, for example, if you were clustering books, you might find a book about Steve Jobs and a book 114 00:08:04,900 --> 00:08:06,800 about Elon Musk in the same cluster. 115 00:08:07,360 --> 00:08:10,660 This cluster is probably about tech companies in Silicon Valley. 116 00:08:12,640 --> 00:08:17,620 But you don't need a label in your data to tell you that you can discover it yourself by looking at 117 00:08:17,920 --> 00:08:19,840 how the data naturally groups together. 118 00:08:21,190 --> 00:08:26,440 Now, as I mentioned earlier, I think once you learn about both supervised and unsupervised learning, 119 00:08:26,710 --> 00:08:29,290 you'll be ready to jump into reinforcement learning. 120 00:08:30,410 --> 00:08:35,780 I haven't made cluster analysis a prerequisite to reinforcement learning, since none of the material 121 00:08:35,780 --> 00:08:36,920 depends on this course. 122 00:08:37,250 --> 00:08:42,050 But it's good to know about these techniques so that you have a more mature and experienced view on 123 00:08:42,050 --> 00:08:42,740 machine learning. 124 00:08:46,320 --> 00:08:50,670 What is sort of a sequel to Cluster Analysis is hidden Markov models. 125 00:08:51,880 --> 00:08:55,330 The reason might not be clear at first, so let me give you two reasons. 126 00:08:56,500 --> 00:09:02,500 Number one, they are both unsupervised machine learning models, just that ACMS is harder, so it's 127 00:09:02,500 --> 00:09:04,540 natural to learn about clustering first. 128 00:09:05,020 --> 00:09:09,040 Clustering is also about static data, whereas ECMS are about sequences. 129 00:09:09,400 --> 00:09:11,950 So it's similar to the process we did in deep learning. 130 00:09:11,950 --> 00:09:16,180 We looked at static data like images, then sequential data like text. 131 00:09:17,760 --> 00:09:23,400 Reason number two, in cluster analysis, we learn about a technique called the Gaussian mixture model, 132 00:09:23,430 --> 00:09:25,800 which we make use of in the same course. 133 00:09:26,580 --> 00:09:31,910 One key point is they both learn by using the expectation maximization algorithm. 134 00:09:32,490 --> 00:09:35,640 So it's good to first see the algorithm on a simple model. 135 00:09:35,640 --> 00:09:41,610 And then when you see them again on a more complicated model like the human, it won't be as intimidating. 136 00:09:43,130 --> 00:09:49,400 One key concept you learn in acronyms is the mark of assumption that just means the current state depends 137 00:09:49,400 --> 00:09:52,700 only on the previous state, but not any states before it. 138 00:09:53,480 --> 00:09:57,410 This is a simplifying assumption that usually makes the math easier to work with. 139 00:09:58,280 --> 00:10:03,140 You will also notice that we encountered the mark of assumption and reinforcement learning, however, 140 00:10:03,410 --> 00:10:05,390 it's not too hard to learn it from scratch. 141 00:10:05,660 --> 00:10:11,060 And so for that reason, I do not consider ECMS to be a prerequisite to reinforcement learning. 142 00:10:11,690 --> 00:10:14,630 The mark of assumption is really the only thing they have in common. 143 00:10:15,590 --> 00:10:22,400 There is also a slight connection between cluster analysis and unsupervised learning, so I'm not going 144 00:10:22,400 --> 00:10:28,670 to draw the link right now, but I sometimes consider this to be unsupervised machining part one and 145 00:10:28,940 --> 00:10:31,400 this to be unsupervised machine learning, part two. 146 00:10:33,420 --> 00:10:38,340 We also see that HMS feeds into the answer, of course, which is about deep learning. 147 00:10:38,580 --> 00:10:39,660 So why might that be? 148 00:10:40,560 --> 00:10:46,140 This is, of course, because both these models are models that can learn about sequences, in particular 149 00:10:46,590 --> 00:10:49,410 in both these courses with model text as sequences. 150 00:10:50,040 --> 00:10:56,190 But whereas the HMS makes use of the mark of assumption, he Arnon does not and hence the Arnon is a 151 00:10:56,190 --> 00:10:57,300 more powerful model. 152 00:10:58,270 --> 00:11:03,280 And so this just goes along with the main theme that we always go from simple basic models to more complex 153 00:11:03,280 --> 00:11:03,790 models. 154 00:11:04,390 --> 00:11:06,730 This is also something you should do in your work as well. 155 00:11:07,330 --> 00:11:11,560 If you start with a simple model, you often find that it is faster and more robust. 156 00:11:11,920 --> 00:11:17,410 Complex models sometimes break down, but they are also more difficult to implement and might not even 157 00:11:17,410 --> 00:11:18,310 be fast enough. 158 00:11:18,670 --> 00:11:20,650 Of course, that's just a generalization. 159 00:11:20,660 --> 00:11:25,300 So you always want to analyze engineering tradeoffs individually for every problem you have. 160 00:11:27,350 --> 00:11:31,970 Now, there is one last link in this part of the graph here that I want to explain, and that's the 161 00:11:31,970 --> 00:11:34,780 first A.P. course you can see there. 162 00:11:34,790 --> 00:11:38,840 It depends on supervised machine learning and feeds into deep enthalpy. 163 00:11:40,220 --> 00:11:46,030 The main purpose of this basic A.P. course is to apply basic machine learning models to text. 164 00:11:46,520 --> 00:11:51,470 So that's why I supervised machine learning comes before it is because that this course was all about 165 00:11:51,680 --> 00:11:53,210 basic machine learning models. 166 00:11:54,110 --> 00:12:00,050 The important skill for NLP was not the implementation of those models, but rather a bigger picture 167 00:12:00,050 --> 00:12:02,500 perspective on how machine learning is used. 168 00:12:03,140 --> 00:12:05,650 What is the interface between the data in the model? 169 00:12:06,080 --> 00:12:07,190 What does the model do? 170 00:12:07,820 --> 00:12:09,400 What is its input and output? 171 00:12:09,560 --> 00:12:10,930 How is the output interpreted? 172 00:12:11,660 --> 00:12:14,780 And so we take those principles and we apply them to text. 173 00:12:15,320 --> 00:12:20,570 In this way we can see that text can be created in such a way that you don't have to think about it 174 00:12:20,810 --> 00:12:23,030 any differently than any other data. 175 00:12:24,310 --> 00:12:29,920 This reinforces the principle that all data is the same, the machine learning model doesn't care what 176 00:12:29,920 --> 00:12:32,790 your data is, all it sees is a table of numbers. 177 00:12:32,800 --> 00:12:36,750 It doesn't care if it's text or images or radar signals from space. 178 00:12:37,270 --> 00:12:41,170 The model just does what it was designed to do on the numbers that you give it. 179 00:12:41,770 --> 00:12:46,930 So this course gives you a high level systems perspective on working with machine learning models in 180 00:12:46,930 --> 00:12:47,470 text. 181 00:12:48,280 --> 00:12:54,430 This easy A.P. course also feeds into deep entropy, which is of course not so easy because it depends 182 00:12:54,430 --> 00:12:56,140 on a lot of background in deep learning. 183 00:12:57,170 --> 00:13:03,260 One of the main questions I get in the NLB course is how do I improve the results of these basic models? 184 00:13:03,830 --> 00:13:08,300 And a lot of the time the answer to that is, well, you have to use a more complex model. 185 00:13:08,310 --> 00:13:12,620 But of course, that necessitates learning how that complex model works. 186 00:13:13,460 --> 00:13:19,460 And Deep in OP is an example of that because we learn a state of the art method for sentiment analysis, 187 00:13:19,460 --> 00:13:22,670 whereas in SCMP we used on the linear model. 188 00:13:23,330 --> 00:13:28,580 So it's important to understand that while, yes, it's possible to improve the predictive ability on 189 00:13:28,580 --> 00:13:33,130 simple basic models, as you can see, it's not always an easy path to get there. 190 00:13:33,590 --> 00:13:35,540 So you have to make sure you're prepared. 191 00:13:36,200 --> 00:13:40,200 Case in point, just look at all the time spent just to get to deep. 192 00:13:40,550 --> 00:13:41,810 It's not an easy task. 193 00:13:43,930 --> 00:13:50,260 Let's now scroll over to Gan's and recreational auto encoders, just like how deep reinforcement learning 194 00:13:50,260 --> 00:13:55,550 is not related to deep in OP, this isn't really related to deep reinforcement learning either. 195 00:13:56,140 --> 00:13:59,650 This is deep learning Part eight by order of creation only. 196 00:14:00,160 --> 00:14:06,310 It is the spiritual sequel to Unsupervised Deep Learning, which was Deep Learning in Python Part four. 197 00:14:07,590 --> 00:14:14,730 OK, so just to reiterate, this is part six, part seven and part eight by order only, but they are 198 00:14:14,730 --> 00:14:20,130 not related to each other conceptually, although it's always nice to know all these things because 199 00:14:20,130 --> 00:14:22,840 more context makes future things easier to learn. 200 00:14:24,230 --> 00:14:32,710 So the reason this is linked to this is because gangs and recreational auto encoders are also unsupervised 201 00:14:32,710 --> 00:14:38,720 deep learning models, but whereas unsupervised deep learning was all about how to improve supervised 202 00:14:38,720 --> 00:14:44,720 learning gains and variation, auto encoders don't have any direct benefit to supervised learning at 203 00:14:44,720 --> 00:14:48,620 all, although we do make use of supervised learning within the course. 204 00:14:49,460 --> 00:14:52,640 In this course, the focus is on generating images. 205 00:14:53,420 --> 00:14:58,940 We've seen that Gan's can create photo realistic images based on a dual neural network system. 206 00:14:59,630 --> 00:15:03,950 That's pretty cool because before Gan's we didn't have any kind of machine learning model that could 207 00:15:03,950 --> 00:15:05,960 generate real looking images. 208 00:15:06,680 --> 00:15:12,800 Nowadays, Gan's are able to generate high resolution, high quality images of people that you can't 209 00:15:12,800 --> 00:15:14,350 even tell are not real people. 210 00:15:14,810 --> 00:15:18,140 It certainly makes the idea of the Matrix seem very possible. 211 00:15:19,800 --> 00:15:22,110 All right, so I hope you found this lecture helpful. 212 00:15:22,620 --> 00:15:27,030 We saw that these courses are related to each other in some pretty complicated ways. 213 00:15:27,840 --> 00:15:30,300 Learning machine learning is not exactly linear. 214 00:15:30,750 --> 00:15:34,140 Sometimes you have to take one course before you can take the next. 215 00:15:34,710 --> 00:15:39,930 Sometimes one course might just be related to another course by some key concept, but maybe in one 216 00:15:39,930 --> 00:15:42,150 context, it's a lot easier to understand. 217 00:15:42,900 --> 00:15:48,600 So remember to keep in mind that these arrows did not all indicate strong prerequisites, or rather 218 00:15:48,600 --> 00:15:51,440 there's just a relationship between the two courses. 219 00:15:52,110 --> 00:15:57,030 I hope that this lecture explained any nuances between what is a prerequisite and what is not. 220 00:15:57,360 --> 00:15:59,340 And I hope I did a good job of answering. 221 00:15:59,670 --> 00:16:02,160 Which order should you take these courses in and why?