1 00:00:02,190 --> 00:00:04,440 Everyone, and welcome back to this class. 2 00:00:06,410 --> 00:00:10,640 This lecture is about how to write code independently when you're doing machine learning. 3 00:00:11,840 --> 00:00:16,400 You always hear me say that the first thing you should try to do when you learn an algorithm is to go 4 00:00:16,400 --> 00:00:20,150 and write it up and code by yourself before looking at someone else's code. 5 00:00:20,990 --> 00:00:25,040 For some students who are new to machine learning, it's not quite clear how to do that. 6 00:00:25,700 --> 00:00:30,950 So this lecture is going to explain the thought process you should go through in hopes that it encourages 7 00:00:30,950 --> 00:00:33,130 more people to try to code by themselves. 8 00:00:36,530 --> 00:00:41,350 I like to encapsulate this in the phrase, When it's time to code, you must code. 9 00:00:41,990 --> 00:00:46,820 So remember this whenever you see some code that demonstrates the concepts we learned in this course, 10 00:00:47,420 --> 00:00:53,030 even if you don't fully understand what's going on, coding and copying examples from others helps you 11 00:00:53,030 --> 00:00:54,050 build muscle memory. 12 00:00:58,460 --> 00:01:02,900 Sometimes people act like they are going to learn tennis by reading a book about tennis. 13 00:01:03,560 --> 00:01:05,790 This is, of course, obviously not possible. 14 00:01:06,680 --> 00:01:11,300 You need to learn some things about tennis, which are pure fact and reside in your brain as conscious 15 00:01:11,300 --> 00:01:11,840 facts. 16 00:01:12,140 --> 00:01:15,520 But then you have to go out and practice those concepts on the tennis court. 17 00:01:16,310 --> 00:01:21,440 Once you've exhausted and mastered everything you know so far, then you can learn new techniques. 18 00:01:21,980 --> 00:01:26,240 So you have this continuous cycle of learning new techniques and then practicing them. 19 00:01:29,640 --> 00:01:36,630 Why is this because learning the technique as pure fact doesn't mean you truly understand it, practicing 20 00:01:36,630 --> 00:01:40,050 the technique helps you think about the technique from a different perspective. 21 00:01:40,560 --> 00:01:46,200 Practicing again and again leads to new understanding, to learning in the subconscious, which is muscle 22 00:01:46,200 --> 00:01:46,580 memory. 23 00:01:47,100 --> 00:01:48,560 You must go through the cycle. 24 00:01:49,440 --> 00:01:54,360 But some students naively think they're going to read an entire book about tennis and then be a tennis 25 00:01:54,360 --> 00:01:56,690 master the first time they try playing. 26 00:01:57,330 --> 00:02:02,820 This is very common, especially when all you really need to do is sit there and watch a video and no 27 00:02:02,820 --> 00:02:03,990 one can force you to type. 28 00:02:04,710 --> 00:02:06,820 A lot of people don't even think to try. 29 00:02:07,740 --> 00:02:09,000 So just remember this. 30 00:02:09,000 --> 00:02:13,110 Whenever you see code, when it's time to code, you must code. 31 00:02:16,630 --> 00:02:22,180 First, we are going to talk a little bit about why you want to code buy yourself a lot of the time 32 00:02:22,180 --> 00:02:25,360 what an algorithm does or how it works isn't quite clear. 33 00:02:25,360 --> 00:02:30,160 The first time you learn about it, every individual will have their own unique background. 34 00:02:30,490 --> 00:02:35,440 So they might be familiar with the patterns and know exactly what to do while some individuals might 35 00:02:35,440 --> 00:02:37,120 have questions that no one else has. 36 00:02:37,900 --> 00:02:43,210 Sometimes you gloss over details because you assume you know what's going on when really there are important 37 00:02:43,210 --> 00:02:44,440 things you're not considering. 38 00:02:44,980 --> 00:02:47,920 So this is how trying to code by yourself can help you. 39 00:02:48,760 --> 00:02:52,070 Coding by yourself forces you to think about each and every detail. 40 00:02:52,480 --> 00:02:54,460 It forces you to think line by line. 41 00:02:55,810 --> 00:03:00,520 You have to be familiar with the data types and the shapes of all your variables, and they all have 42 00:03:00,520 --> 00:03:03,550 to fit together properly, kind of like Lego blocks. 43 00:03:07,400 --> 00:03:13,130 As an example, you know that in order to do element wise matrix additions, both matrices have to have 44 00:03:13,130 --> 00:03:14,330 the exact same shape. 45 00:03:15,050 --> 00:03:20,030 So if you find that they are not the same shape, then one of your previous assumptions was incorrect. 46 00:03:20,990 --> 00:03:22,730 So you should go back and correct it. 47 00:03:23,780 --> 00:03:28,240 Lego blocks have to follow a specific set of rules in order to fit together properly. 48 00:03:29,000 --> 00:03:33,900 If you make incorrect assumptions about how Lego blocks fit together and you try to join them. 49 00:03:33,920 --> 00:03:37,840 It's not going to work without trying to build things by yourself. 50 00:03:38,120 --> 00:03:40,040 You will never discover these details. 51 00:03:43,840 --> 00:03:46,310 Let's now talk about how to code by yourself. 52 00:03:46,960 --> 00:03:49,510 Consider the supervised machine learning scenario. 53 00:03:50,230 --> 00:03:55,150 We know that we're going to have some inputs and some targets, and we want to try to make predictions 54 00:03:55,360 --> 00:03:58,000 from the inputs that are very close to the targets. 55 00:03:58,750 --> 00:04:01,390 We typically call the inputs X and the targets. 56 00:04:01,390 --> 00:04:01,780 Why? 57 00:04:02,440 --> 00:04:07,000 Sometimes we call the targets T, but for the purpose of this lecture, we'll call the targets Y. 58 00:04:09,910 --> 00:04:16,360 Now, this next point is a key point, it doesn't actually matter what X is and what why is it doesn't 59 00:04:16,360 --> 00:04:21,730 matter if you're looking at an e-commerce data set and the columns of X might be time on site, time 60 00:04:21,730 --> 00:04:24,640 of day, how many pages the user looked at and so on. 61 00:04:25,210 --> 00:04:30,730 You could just as well be looking at a data set of x ray images where each column is the pixel intensity 62 00:04:30,730 --> 00:04:31,360 of an image. 63 00:04:31,990 --> 00:04:33,190 This is the key point. 64 00:04:33,460 --> 00:04:37,650 We say all data is the same when we do linear regression. 65 00:04:37,660 --> 00:04:41,650 We don't have different kinds of linear regression for e commerce and x ray images. 66 00:04:42,190 --> 00:04:45,930 Linear regression is the same algorithm no matter what your data set is. 67 00:04:46,300 --> 00:04:48,310 So we say all data is the same. 68 00:04:49,350 --> 00:04:54,810 Theoretically, I could give you a data set of axes and Y's and you could train a classification algorithm 69 00:04:54,810 --> 00:05:00,720 on it without me even telling you what the Xs and Ys mean, you should get very comfortable with this 70 00:05:00,720 --> 00:05:01,200 idea. 71 00:05:04,470 --> 00:05:10,020 Sometimes people say, well, this isn't practical because I want to do concrete examples, but that's 72 00:05:10,020 --> 00:05:11,940 because they're not thinking intelligently. 73 00:05:12,540 --> 00:05:17,400 In fact, this is the most practical thing we can do because it means everything we learn. 74 00:05:17,610 --> 00:05:20,190 We can apply it to any data set that exists. 75 00:05:20,820 --> 00:05:26,280 It means that I can train a model on an e-commerce data set, but I can also train a model on an online 76 00:05:26,280 --> 00:05:29,320 advertising data set without learning anything new at all. 77 00:05:29,850 --> 00:05:32,330 It's truly the greatest form of lazy programming. 78 00:05:32,760 --> 00:05:35,460 Learn something once and apply it to every industry. 79 00:05:39,090 --> 00:05:45,090 One great consequence of all data is the same is that there is an unlimited amount of practice opportunity 80 00:05:45,090 --> 00:05:45,470 for you. 81 00:05:46,230 --> 00:05:51,690 You can download data sets from Kaggle, from Google, from Wikipedia, from Amazon or from anywhere 82 00:05:51,690 --> 00:05:53,850 else and try the algorithms you learned. 83 00:05:54,630 --> 00:06:00,560 What this means is let's say you have a data set you care about more than something we do in Class C 84 00:06:00,600 --> 00:06:01,040 and class. 85 00:06:01,050 --> 00:06:05,940 We have to use data sets that everybody can understand, for example, text and images. 86 00:06:06,210 --> 00:06:08,460 Everybody should know what text and images are. 87 00:06:09,240 --> 00:06:12,450 Text and images are just fundamental data types on the web. 88 00:06:12,990 --> 00:06:17,820 You probably don't even think about the fact that text and images are data because it's so trivial. 89 00:06:18,780 --> 00:06:23,100 In any case, we work with text and images a lot because everybody understands them. 90 00:06:23,850 --> 00:06:26,820 But let's say you're a biologist and you know about DNA. 91 00:06:27,060 --> 00:06:29,720 You find DNA and genomics really interesting. 92 00:06:30,360 --> 00:06:32,600 So you want to use the algorithms we learn on that. 93 00:06:33,210 --> 00:06:33,900 Well, that's great. 94 00:06:33,900 --> 00:06:35,030 And you should totally do it. 95 00:06:35,640 --> 00:06:37,410 Remember, all data is the same. 96 00:06:37,680 --> 00:06:43,530 So all you need to do is convert your data into the appropriate Xs and whys to plug into our machine 97 00:06:43,530 --> 00:06:44,440 learning algorithms. 98 00:06:45,690 --> 00:06:52,080 The downside is we most likely won't talk about DNA in class unless it's a very simple example because 99 00:06:52,080 --> 00:06:54,720 most computer scientists are not biologists too. 100 00:06:55,440 --> 00:06:57,300 So they don't know what DNA is. 101 00:06:57,300 --> 00:06:58,970 They don't understand the specifics. 102 00:06:59,430 --> 00:07:01,600 So an example wouldn't make much sense to them. 103 00:07:02,310 --> 00:07:05,520 This is opposed to text and images, which makes sense to everybody. 104 00:07:06,360 --> 00:07:12,090 Same goes with any other specialized field like finance, computer networking, cosmology and so on. 105 00:07:16,860 --> 00:07:21,840 So now we can start off from the right place to talk about supervised learning, we have our exes and 106 00:07:21,840 --> 00:07:26,060 we have our wise, what do we do with them in supervised learning? 107 00:07:26,070 --> 00:07:28,290 We know that there are two main things we want to do. 108 00:07:28,960 --> 00:07:33,090 We want to do training and we want to do prediction and psyche to learn. 109 00:07:33,090 --> 00:07:34,910 All the models have the same API. 110 00:07:35,310 --> 00:07:41,070 It doesn't matter if you're doing logistic regression or decision trees or random forest unsupervised 111 00:07:41,070 --> 00:07:45,240 models in Saikat Learn have the same two functions fit and predict. 112 00:07:45,840 --> 00:07:48,570 Fitting a model is just a synonym for training a model. 113 00:07:49,080 --> 00:07:53,070 If you don't believe me, you can go look at the psychic documentation yourself. 114 00:07:56,590 --> 00:08:01,690 When you're learning about supervised machine learning, what you're really learning is what goes inside 115 00:08:01,690 --> 00:08:05,890 these two functions, what are the parameters and how are those parameters learned? 116 00:08:06,580 --> 00:08:07,990 That's really what I'll supervise. 117 00:08:07,990 --> 00:08:10,810 Machine learning is filling in these two functions. 118 00:08:11,470 --> 00:08:16,300 Taking this perspective is going to give your code structure and it's going to make things much easier 119 00:08:16,510 --> 00:08:17,890 to visualize in your head. 120 00:08:20,970 --> 00:08:27,570 Let's now do a simple example to solidify this idea and this example, I'm going to give you an algorithm 121 00:08:27,570 --> 00:08:29,300 and you will implement it in code. 122 00:08:30,030 --> 00:08:33,290 There are two key points I want to make before I give you the algorithm. 123 00:08:33,990 --> 00:08:37,710 First, I'm not going to give you much intuition about how it works. 124 00:08:38,280 --> 00:08:41,710 Second, I'm not going to derive the theory behind why it works. 125 00:08:42,450 --> 00:08:47,640 The reason I want to mention these two key points is that you should realize you don't need these two 126 00:08:47,640 --> 00:08:51,650 pieces of information in order to translate pseudocode into code. 127 00:08:52,290 --> 00:08:58,560 A lot of time, these three concepts, intuition, theory and implementation work to reinforce each 128 00:08:58,560 --> 00:08:58,860 other. 129 00:08:59,670 --> 00:09:05,310 The reason I'm focusing on implementation right now is because it's the one people most often miss or 130 00:09:05,310 --> 00:09:10,200 think they don't need at all, when really it's an extremely important part of the learning process. 131 00:09:13,540 --> 00:09:19,630 OK, so the pseudocode is as follows in my predict function, as you already know, I'm going to take 132 00:09:19,630 --> 00:09:24,850 in some input data X my predictions will be Y hat equals X times W. 133 00:09:25,690 --> 00:09:28,420 You might already recognize this as linear regression. 134 00:09:29,050 --> 00:09:34,240 However, forget what you know about linear regression and pretend that this is just some formula I 135 00:09:34,240 --> 00:09:34,720 gave you. 136 00:09:35,500 --> 00:09:38,950 What I can tell you is that we are creating some kind of regression model. 137 00:09:39,650 --> 00:09:44,230 It should be clear from this equation that the parameters of the model are contained in this vector 138 00:09:44,230 --> 00:09:44,650 of weights. 139 00:09:44,650 --> 00:09:51,010 W I would probably also tell you that W is a vector that's the same size as the number of feature vectors 140 00:09:51,010 --> 00:09:51,580 in X. 141 00:09:52,210 --> 00:09:58,360 Of course this must be true in order for the Matrix multiplied to be valid, this axis kind of a sanity 142 00:09:58,360 --> 00:10:01,120 check so you can make sure everything is making sense. 143 00:10:04,450 --> 00:10:10,720 In my fit function, I want to perform this loop some number of times, since we know from the previous 144 00:10:10,720 --> 00:10:15,970 slide that the parameters of the model are contained and w then it should be clear that inside the fit 145 00:10:15,970 --> 00:10:18,400 function, what we want to do is update w. 146 00:10:19,060 --> 00:10:21,460 Not surprisingly, that's what's happening here. 147 00:10:21,640 --> 00:10:23,260 Another good sanity check. 148 00:10:24,180 --> 00:10:29,310 Now, because you're already familiar with linear regression, you probably already recognize this as 149 00:10:29,310 --> 00:10:30,120 gradient descent. 150 00:10:30,690 --> 00:10:34,710 Again, I want you to pretend that you don't know that, but what do you know? 151 00:10:38,530 --> 00:10:43,150 You know, that iterative of algorithms are common in machine learning, and that gradient descent is 152 00:10:43,150 --> 00:10:48,880 among the most common, you could probably infer that this is some form of gradient descent, but you 153 00:10:48,880 --> 00:10:51,560 don't need to know what it is in order to write it in code. 154 00:10:52,330 --> 00:10:53,280 What else do we know? 155 00:10:54,130 --> 00:10:59,500 You know what a for loop is and you know how to write one in Python, you know that axis and by the 156 00:10:59,500 --> 00:11:04,540 Matrix containing the input data and that Y is an inside vector containing the targets. 157 00:11:05,470 --> 00:11:11,290 You know, that white hat is an inside vector containing the predictions, you know, that W is a disease 158 00:11:11,290 --> 00:11:16,810 vector containing the model parameters and you know how to add, subtract and multiply these objects 159 00:11:17,260 --> 00:11:20,450 so we can write this algorithm without even knowing how it works. 160 00:11:21,250 --> 00:11:24,690 Of course, a lot about how it works can just be inferred from this formula. 161 00:11:28,740 --> 00:11:33,420 Here is what you don't know, you don't know what the number of times the loop is supposed to run, 162 00:11:33,870 --> 00:11:35,530 you don't know, Ayda, the learning rate. 163 00:11:36,090 --> 00:11:37,430 This is completely fine. 164 00:11:37,980 --> 00:11:41,010 You can't let this lack of knowledge stop you in your tracks. 165 00:11:41,760 --> 00:11:46,680 The truth of the matter is they're going to be situations where you're not told the exact numbers to 166 00:11:46,680 --> 00:11:47,240 plug in. 167 00:11:47,760 --> 00:11:51,570 In many cases, the answer is going to be it depends on the problem. 168 00:11:52,410 --> 00:11:58,470 The important thing is you have to get used to the idea of trial and error and you have to make educated 169 00:11:58,470 --> 00:12:00,570 guesses based on what you already know. 170 00:12:04,100 --> 00:12:09,950 So what can you do when you know that the learning rate is supposed to be a small number, how small? 171 00:12:10,130 --> 00:12:11,520 It depends on the problem. 172 00:12:12,320 --> 00:12:15,320 Typically, it's a number less than one like zero point one. 173 00:12:15,830 --> 00:12:17,930 If that doesn't work, you might try lowering it. 174 00:12:18,500 --> 00:12:20,650 Typically, we lower it by factors of 10. 175 00:12:21,260 --> 00:12:26,120 So, for example, the next learning rate we would try is ten to the minus two tenths of a minus three 176 00:12:26,120 --> 00:12:26,810 and so on. 177 00:12:30,950 --> 00:12:36,320 Typically in machine learning, whether you're looking at a supervised or unsupervised algorithm, there 178 00:12:36,320 --> 00:12:39,940 is an associated cost or objective function that you're trying to minimize. 179 00:12:40,730 --> 00:12:45,260 Let's suppose I tell you it's squarer, which is typically what we would use for regression. 180 00:12:46,010 --> 00:12:50,380 Well, now you have a way to choose the number of iterations of the loop and the learning rate. 181 00:12:50,870 --> 00:12:51,860 How do we do this? 182 00:12:56,140 --> 00:13:02,380 So inside your fit function, you would plot the cost as a function of iteration, I always recommend 183 00:13:02,380 --> 00:13:02,890 doing this. 184 00:13:02,890 --> 00:13:08,110 No matter what algorithm you're looking at in general, you always want the cost to converge. 185 00:13:08,770 --> 00:13:13,990 The pattern you typically see is that there is a steep drop at the beginning and it slowly flattens 186 00:13:13,990 --> 00:13:19,540 out as the number of iterations increases to choose the number of iterations you want to stop. 187 00:13:19,540 --> 00:13:25,350 When the curve is sufficiently flat, you might recognize this as a diminishing returns scenario. 188 00:13:26,410 --> 00:13:32,230 The returns diminish because at the beginning, we reduce the cost by a lot in just a few iterations 189 00:13:32,680 --> 00:13:36,820 at the end we can do a lot of iterations, but we reduce the costs by almost nothing. 190 00:13:39,870 --> 00:13:45,930 An alternative is you could use a separate validation set and stop when the validation cost increases, 191 00:13:46,050 --> 00:13:48,120 but that is not the focus of this lecture. 192 00:13:48,690 --> 00:13:54,210 Unlike the training cost, the validation cost is not guaranteed to decrease at every round because 193 00:13:54,210 --> 00:13:56,910 that is not what we are minimizing with respect to. 194 00:14:00,710 --> 00:14:06,440 You can also use the same plot to help improve your learning rate if the cost explodes or becomes not 195 00:14:06,440 --> 00:14:11,670 a number, your learning rate is probably too high if your cost converges to slowly. 196 00:14:11,690 --> 00:14:13,640 You can try increasing your learning rate. 197 00:14:17,890 --> 00:14:19,950 So what does your final code look like? 198 00:14:21,790 --> 00:14:26,740 Well, we know we want to create a class with at least that predict and fit functions, you don't have 199 00:14:26,740 --> 00:14:31,480 to use a class, but I find that it encapsulates the code nicely and provides useful structure. 200 00:14:33,120 --> 00:14:37,910 Notice that once you have the algorithm in math, there isn't much work to convert it into numbered 201 00:14:37,920 --> 00:14:38,260 code. 202 00:14:38,820 --> 00:14:40,410 You know how to do matrix model. 203 00:14:41,010 --> 00:14:43,100 You know how to do matrix multiplication. 204 00:14:43,380 --> 00:14:45,660 You know how to add you know how to subtract. 205 00:14:46,180 --> 00:14:49,530 These are basic arithmetic operations that I'm sure you know already. 206 00:14:50,280 --> 00:14:51,600 If you don't know how to do these. 207 00:14:51,600 --> 00:14:56,460 And then by then I have a completely free number because on the number I stack that you can take. 208 00:15:00,110 --> 00:15:04,860 Here is the second part of the code where we actually create an instance of the class and then use it. 209 00:15:05,480 --> 00:15:09,260 I've also included the squared error cost function just for completion sake. 210 00:15:14,380 --> 00:15:19,100 What's cool about this template is that it doesn't really change from one algorithm to the next. 211 00:15:19,600 --> 00:15:24,820 The structure is basically always going to be this way, at least for supervised learning and unsupervised 212 00:15:24,820 --> 00:15:25,200 learning. 213 00:15:26,020 --> 00:15:30,490 So as we discussed before, it doesn't matter what algorithm you're implementing, whether it be linear 214 00:15:30,490 --> 00:15:35,830 regression, logistic regression, a neural network and so on, they always have the predictive function 215 00:15:35,830 --> 00:15:42,040 and the fit function and learning what goes inside these functions is equivalent to learning the algorithm. 216 00:15:43,580 --> 00:15:49,100 So, in fact, one way to begin coding is to start with just this boilerplate code and then fill in 217 00:15:49,100 --> 00:15:49,730 the blanks.