1 00:00:02,220 --> 00:00:04,480 Hey, everyone, and welcome back to this class. 2 00:00:06,450 --> 00:00:10,620 This lecture is about how to write code independently when you're doing machine learning. 3 00:00:11,910 --> 00:00:16,410 You always hear me say that the first thing you should try to do when you learn an algorithm is to go 4 00:00:16,410 --> 00:00:20,130 and write it up and code by yourself before looking at someone else's code. 5 00:00:21,060 --> 00:00:25,020 For some students who are new to machine learning, it's not quite clear how to do that. 6 00:00:25,770 --> 00:00:30,960 So this lecture is going to explain the thought process you should go through in hopes that it encourages 7 00:00:30,960 --> 00:00:33,120 more people to try to code by themselves. 8 00:00:36,660 --> 00:00:41,370 I like to encapsulate this in the phrase, When it's time to code, you must code. 9 00:00:42,060 --> 00:00:46,830 So remember this whenever you see some code that demonstrates the concepts we learned in this course, 10 00:00:47,520 --> 00:00:53,040 even if you don't fully understand what's going on, coding and copying examples from others helps you 11 00:00:53,040 --> 00:00:54,060 build muscle memory. 12 00:00:58,530 --> 00:01:02,970 Sometimes people act like they are going to learn tennis by reading a book about tennis. 13 00:01:03,630 --> 00:01:05,820 This is, of course, obviously not possible. 14 00:01:06,780 --> 00:01:11,280 You need to learn some things about tennis, which are pure fact and reside in your brain as conscious 15 00:01:11,280 --> 00:01:11,850 facts. 16 00:01:12,210 --> 00:01:15,510 But then you have to go out and practice those concepts on the tennis court. 17 00:01:16,380 --> 00:01:21,450 Once you've exhausted and mastered everything you know so far, then you can learn new techniques. 18 00:01:22,050 --> 00:01:26,270 So you have this continuous cycle of learning new techniques and then practicing them. 19 00:01:29,730 --> 00:01:30,510 Why is this? 20 00:01:31,260 --> 00:01:35,400 Because learning the technique as pure fact, does it mean you truly understand it? 21 00:01:36,090 --> 00:01:40,050 Practicing the technique helps you think about the technique from a different perspective. 22 00:01:40,650 --> 00:01:46,200 Practicing again and again leads to new understanding, to learning in the subconscious, which is muscle 23 00:01:46,200 --> 00:01:46,590 memory. 24 00:01:47,190 --> 00:01:48,600 You must go through the cycle. 25 00:01:49,500 --> 00:01:54,360 But some students naively think they are going to read an entire book about tennis and then be a tennis 26 00:01:54,360 --> 00:01:54,990 master. 27 00:01:55,260 --> 00:02:01,080 The first time they try playing this is very common, especially when all you really need to do is sit 28 00:02:01,080 --> 00:02:04,050 there and watch a video and no one can force you to type. 29 00:02:04,830 --> 00:02:06,810 A lot of people don't even think to try. 30 00:02:07,830 --> 00:02:09,000 So just remember this. 31 00:02:09,000 --> 00:02:13,110 Whenever you see code, when it's time to code, you must code. 32 00:02:16,670 --> 00:02:22,400 First, we are going to talk a little bit about why you want to code by yourself a lot of the time what 33 00:02:22,400 --> 00:02:26,630 an algorithm does or how it works isn't quite clear the first time you learn about it. 34 00:02:27,530 --> 00:02:30,160 Every individual will have their own unique background. 35 00:02:30,560 --> 00:02:33,740 So they might be familiar with the patterns and know exactly what to do. 36 00:02:34,250 --> 00:02:37,160 While some individuals might have questions that no one else has. 37 00:02:37,970 --> 00:02:43,220 Sometimes you gloss over details because you assume you know what's going on when really there are important 38 00:02:43,220 --> 00:02:44,480 things you're not considering. 39 00:02:45,050 --> 00:02:47,930 So this is how trying to code by yourself can help you. 40 00:02:48,830 --> 00:02:49,740 Coding by yourself. 41 00:02:49,760 --> 00:02:52,070 Forces you to think about each and every detail. 42 00:02:52,580 --> 00:02:54,430 It forces you to think line by line. 43 00:02:55,910 --> 00:03:00,530 You have to be familiar with the data types and the shapes of all your variables, and they all have 44 00:03:00,530 --> 00:03:01,820 to fit together properly. 45 00:03:02,180 --> 00:03:03,590 Kind of like Lego blocks. 46 00:03:07,440 --> 00:03:13,140 As an example, you know that in order to do element wise matrix additions, both matrices have to have 47 00:03:13,140 --> 00:03:14,370 the exact same shape. 48 00:03:15,120 --> 00:03:20,050 So if you find that they are not the same shape, then one of your previous assumptions was incorrect. 49 00:03:21,060 --> 00:03:22,710 So you should go back and correct it. 50 00:03:23,850 --> 00:03:28,260 Lego blocks have to follow a specific set of rules in order to fit together properly. 51 00:03:29,100 --> 00:03:34,140 If you make incorrect assumptions about how Lego blocks fit together and you try to join them, it's 52 00:03:34,140 --> 00:03:37,830 not going to work without trying to build things by yourself. 53 00:03:38,220 --> 00:03:40,050 You will never discover these details. 54 00:03:43,940 --> 00:03:46,340 Let's now talk about how to code by yourself. 55 00:03:47,030 --> 00:03:49,520 Consider this supervised machine learning scenario. 56 00:03:50,300 --> 00:03:55,130 We know that we're going to have some inputs and some targets, and we want to try to make predictions 57 00:03:55,400 --> 00:03:58,000 from the inputs that are very close to the targets. 58 00:03:58,850 --> 00:04:01,750 We typically call the inputs X and the targets Y. 59 00:04:02,510 --> 00:04:04,100 Sometimes we call the targets T. 60 00:04:04,160 --> 00:04:06,980 But for the purpose of this lecture, we'll call the targets Y. 61 00:04:10,010 --> 00:04:11,780 Now, this next point is a key point. 62 00:04:12,530 --> 00:04:18,140 It doesn't actually matter what X is and what why is it does it matter if you're looking at an E commerce 63 00:04:18,220 --> 00:04:22,190 dataset and the columns of X might be time on site, time of day? 64 00:04:22,460 --> 00:04:24,650 How many pages the user looked at and so on. 65 00:04:25,310 --> 00:04:30,740 You could just as well be looking at a data set of x ray images where each column is the pixel intensity 66 00:04:30,740 --> 00:04:31,400 of an image. 67 00:04:32,060 --> 00:04:33,170 This is the key point. 68 00:04:33,560 --> 00:04:35,300 We say all data is the same. 69 00:04:36,140 --> 00:04:40,790 When we do linear regression, we don't have different kinds of linear regression for e commerce and 70 00:04:40,790 --> 00:04:41,660 x ray images. 71 00:04:42,260 --> 00:04:45,980 Linear regression is the same algorithm no matter what your dataset is. 72 00:04:46,400 --> 00:04:48,350 So we say all data is the same. 73 00:04:49,400 --> 00:04:54,800 Theoretically, I could give you a data set of X's and Y's and you could train a classification algorithm 74 00:04:54,800 --> 00:04:58,490 on it without me even telling you what the Xs and Ys mean. 75 00:04:59,210 --> 00:05:01,190 You should get very comfortable with this idea. 76 00:05:04,550 --> 00:05:10,040 Sometimes people say, well, this isn't practical because I want to do concrete examples, but that's 77 00:05:10,040 --> 00:05:11,990 because they're not thinking intelligently. 78 00:05:12,650 --> 00:05:17,420 In fact, this is the most practical thing we can do because it means everything we learn. 79 00:05:17,690 --> 00:05:20,240 We can apply it to any dataset that exists. 80 00:05:20,960 --> 00:05:26,270 It means that I can train a model on an e-commerce dataset, but I can also train a model on an online 81 00:05:26,270 --> 00:05:29,330 advertising dataset without learning anything new at all. 82 00:05:29,960 --> 00:05:32,360 It's truly the greatest form of lazy programming. 83 00:05:32,840 --> 00:05:35,480 Learn something once and apply it to every industry. 84 00:05:39,190 --> 00:05:45,070 One great consequence of all data is the same is that there is an unlimited amount of practice opportunity 85 00:05:45,070 --> 00:05:45,460 for you. 86 00:05:46,330 --> 00:05:51,700 You can download data sets from Kaggle, from Google, from Wikipedia, from Amazon or from anywhere 87 00:05:51,700 --> 00:05:53,860 else and try the algorithms you learned. 88 00:05:54,700 --> 00:05:57,170 What this means is let's say you have a data set. 89 00:05:57,340 --> 00:06:01,020 You care about more than something we do in Class C and class. 90 00:06:01,030 --> 00:06:03,670 We have to use data sets that everybody can understand. 91 00:06:04,210 --> 00:06:05,980 For example, text and images. 92 00:06:06,310 --> 00:06:08,440 Everybody should know what text and images are. 93 00:06:09,280 --> 00:06:12,460 Text and images are just fundamental data types on the web. 94 00:06:13,090 --> 00:06:17,830 You probably don't even think about the fact that texts and images are data because it's so trivial. 95 00:06:18,910 --> 00:06:23,080 In any case, we work with text and images a lot because everybody understands them. 96 00:06:23,920 --> 00:06:26,800 But let's say you're a biologist and you know about DNA. 97 00:06:27,130 --> 00:06:29,710 You find DNA and genomics really interesting. 98 00:06:30,430 --> 00:06:31,840 So you want to use the algorithms. 99 00:06:31,840 --> 00:06:32,590 We learn on that. 100 00:06:33,280 --> 00:06:33,910 Well, that's great. 101 00:06:33,910 --> 00:06:35,020 And you should totally do it. 102 00:06:35,740 --> 00:06:37,390 Remember, all data is the same. 103 00:06:37,720 --> 00:06:43,810 So all you need to do is convert your data into the appropriate Xs in Ys to plug into our machine learning 104 00:06:43,810 --> 00:06:44,470 algorithms. 105 00:06:45,760 --> 00:06:52,090 The downside is we most likely won't talk about DNA in class unless it's a very simple example because 106 00:06:52,090 --> 00:06:54,660 most computer scientists are not biologists, too. 107 00:06:55,540 --> 00:06:57,220 So they don't know what DNA is. 108 00:06:57,370 --> 00:06:58,990 They don't understand the specifics. 109 00:06:59,530 --> 00:07:01,600 So an example wouldn't make much sense to them. 110 00:07:02,350 --> 00:07:05,530 This is opposed to text and images, which makes sense to everybody. 111 00:07:06,430 --> 00:07:12,120 Same goes with any other specialized field like finance, computer networking, cosmology and so on. 112 00:07:16,930 --> 00:07:20,380 So now we can start from the right place to talk about supervised learning. 113 00:07:20,920 --> 00:07:22,800 We have our exes and we have our wise. 114 00:07:22,930 --> 00:07:26,050 What do we do with them in supervised learning? 115 00:07:26,110 --> 00:07:28,300 We know that there are two main things we want to do. 116 00:07:29,050 --> 00:07:33,070 We want to do training and we want to do prediction and psyche to learn. 117 00:07:33,100 --> 00:07:34,900 All the models have the same API. 118 00:07:35,440 --> 00:07:41,070 It doesn't matter if you're doing logistic regression or decision trees or random forest, all supervised 119 00:07:41,080 --> 00:07:45,250 models in Saikat learn to have the same two functions fit and predict. 120 00:07:45,910 --> 00:07:48,550 Fitting a model is just a synonym for training, a model. 121 00:07:49,180 --> 00:07:53,080 If you don't believe me, you can go look at these Saikat documentation yourself. 122 00:07:56,690 --> 00:08:01,690 When you're learning about supervised machine learning, what you're really learning is what goes inside 123 00:08:01,700 --> 00:08:02,660 these two functions. 124 00:08:03,230 --> 00:08:05,870 What are the parameters and how are those parameters learned? 125 00:08:06,650 --> 00:08:08,000 That's really what I'll supervised. 126 00:08:08,000 --> 00:08:09,740 Machine learning is filling in. 127 00:08:09,790 --> 00:08:15,050 These two functions, taking this perspective is going to give you a code structure and it's going to 128 00:08:15,050 --> 00:08:17,900 make things much easier to visualize in your head. 129 00:08:21,040 --> 00:08:27,580 Let's now do a simple example to solidify this idea and this example, I'm going to give you an algorithm 130 00:08:27,640 --> 00:08:29,320 and you will implement it in code. 131 00:08:30,100 --> 00:08:33,310 There are two key points I want to make before I give you the algorithm. 132 00:08:34,030 --> 00:08:37,750 First, I'm not going to give you much intuition about how it works. 133 00:08:38,380 --> 00:08:41,740 Second, I'm not going to derive the theory behind why it works. 134 00:08:42,520 --> 00:08:47,620 The reason I want to mention these two key points is that you should realize you don't need these two 135 00:08:47,620 --> 00:08:51,640 pieces of information in order to translate pseudocode into code. 136 00:08:52,420 --> 00:08:58,570 A lot of time, these three concepts, intuition, theory and implementation work to reinforce each 137 00:08:58,570 --> 00:08:58,870 other. 138 00:08:59,740 --> 00:09:05,280 The reason I'm focusing on implementation right now is because it's the one people most often miss or 139 00:09:05,290 --> 00:09:10,240 think they don't need at all, when really it's an extremely important part of the learning process. 140 00:09:13,680 --> 00:09:14,000 OK. 141 00:09:14,090 --> 00:09:15,980 So the pseudocode is as follows. 142 00:09:16,490 --> 00:09:18,770 In my predict function, as you already know. 143 00:09:18,890 --> 00:09:24,830 I'm going to take in some input data X my predictions will be Y hat equals X times W. 144 00:09:25,790 --> 00:09:28,430 You might already recognize this as linear regression. 145 00:09:29,120 --> 00:09:34,250 However, forget what you know about linear regression and pretend that this is just some formula I 146 00:09:34,250 --> 00:09:34,730 gave you. 147 00:09:35,570 --> 00:09:38,960 What I can tell you is that we are creating some kind of regression model. 148 00:09:39,770 --> 00:09:44,240 It should be clear from this equation that the parameters of the model are contained in this vector 149 00:09:44,240 --> 00:09:44,620 of weights. 150 00:09:44,630 --> 00:09:51,020 W I would probably also tell you that W is a vector that's the same size as the number of feature vectors 151 00:09:51,020 --> 00:09:51,610 in X. 152 00:09:52,340 --> 00:09:56,060 Of course this must be true in order for the Matrix multiplied to be valid. 153 00:09:56,930 --> 00:09:58,670 This axis kind of a sanity check. 154 00:09:58,700 --> 00:10:01,130 So you can make sure everything is making sense. 155 00:10:04,550 --> 00:10:10,700 In my fit function, I want to perform this loop some number of times, since we know from the previous 156 00:10:10,700 --> 00:10:15,950 slide that the parameters of the model are contained and w then it should be clear that inside the fit 157 00:10:15,950 --> 00:10:18,410 function, what we want to do is update w. 158 00:10:19,160 --> 00:10:21,440 Not surprisingly, that's what's happening here. 159 00:10:21,800 --> 00:10:23,270 Another good sanity check. 160 00:10:24,290 --> 00:10:29,300 Now, because you're already familiar with linear regression, you probably already recognize this as 161 00:10:29,300 --> 00:10:30,110 gradient descent. 162 00:10:30,830 --> 00:10:34,700 Again, I want you to pretend that you don't know that, but what do you know? 163 00:10:38,630 --> 00:10:43,490 You know that iterative algorithms are common in machine learning and that gradient descent is among 164 00:10:43,490 --> 00:10:44,360 the most common. 165 00:10:45,020 --> 00:10:49,820 You could probably infer that this is some form of gradient descent, but you don't need to know what 166 00:10:49,820 --> 00:10:51,590 it is in order to write it in code. 167 00:10:52,430 --> 00:10:53,270 What else do we know? 168 00:10:54,260 --> 00:10:55,520 You know what a for loop is. 169 00:10:55,540 --> 00:10:57,290 And you know how to write one in Python. 170 00:10:58,040 --> 00:11:01,310 You know that X is an end by the Matrix containing the input data. 171 00:11:01,670 --> 00:11:04,550 And that Y is an end size vector containing the targets. 172 00:11:05,570 --> 00:11:08,900 You know that why hat is an inside vector containing the predictions. 173 00:11:09,560 --> 00:11:13,250 You know that W is a disease vector containing the model parameters. 174 00:11:13,790 --> 00:11:19,070 And you know how to add, subtract and multiply these objects so we can write this algorithm without 175 00:11:19,100 --> 00:11:20,480 even knowing how it works. 176 00:11:21,410 --> 00:11:24,650 Of course, a lot about how it works can just be inferred from this formula. 177 00:11:28,790 --> 00:11:30,110 Here is what you don't know. 178 00:11:30,470 --> 00:11:33,410 You don't know T- the number of times the loop is supposed to run. 179 00:11:33,950 --> 00:11:35,540 You don't know, Ayda, the learning rate. 180 00:11:36,170 --> 00:11:37,490 This is completely fine. 181 00:11:38,090 --> 00:11:41,030 You can't let this lack of knowledge stop you in your tracks. 182 00:11:41,810 --> 00:11:46,580 The truth of the matter is there are going to be situations where you're not told the exact numbers 183 00:11:46,580 --> 00:11:47,240 to plug in. 184 00:11:47,900 --> 00:11:51,560 In many cases, the answer is going to be it depends on the problem. 185 00:11:52,430 --> 00:11:58,460 The important thing is you have to get used to the idea of trial and error and you have to make educated 186 00:11:58,460 --> 00:12:00,590 guesses based on what you already know. 187 00:12:04,160 --> 00:12:08,900 So what can you do when you know that the learning rate is supposed to be a small number? 188 00:12:09,320 --> 00:12:09,950 How small? 189 00:12:10,250 --> 00:12:11,540 It depends on the problem. 190 00:12:12,350 --> 00:12:15,350 Typically, it's a number less than one like zero point one. 191 00:12:15,950 --> 00:12:17,900 If that doesn't work, you might try lowering it. 192 00:12:18,560 --> 00:12:20,660 Typically, we lower it by factors of 10. 193 00:12:21,320 --> 00:12:25,050 So, for example, the next learning rate we would try is ten to the minus two. 194 00:12:25,100 --> 00:12:26,810 Ten to the minus three and so on. 195 00:12:30,990 --> 00:12:36,330 Typically in machine learning, whether you're looking at a supervised or unsupervised algorithm, there 196 00:12:36,330 --> 00:12:39,990 is an associated cost or objective function that you're trying to minimize. 197 00:12:40,830 --> 00:12:45,300 Let's suppose I tell you it's square, which is typically what we would use for regression. 198 00:12:46,120 --> 00:12:50,370 Well, now you have a way to choose the number of iterations of the loop and the learning rate. 199 00:12:50,970 --> 00:12:51,870 How do we do this? 200 00:12:56,220 --> 00:13:00,600 So inside your fit function, you would plot the cost as a function of iteration? 201 00:13:01,410 --> 00:13:04,710 I always recommend doing this no matter what algorithm you're looking at. 202 00:13:05,640 --> 00:13:08,130 In general, you always want the cost to converge. 203 00:13:08,850 --> 00:13:14,010 The pattern you typically see is that there is a steep drop at the beginning and it slowly flattens 204 00:13:14,010 --> 00:13:19,770 out as the number of iterations increases to choose the number of iterations you want to stop when the 205 00:13:19,770 --> 00:13:21,300 curve is sufficiently flat. 206 00:13:22,170 --> 00:13:25,320 You might recognize this as a diminishing returns scenario. 207 00:13:26,480 --> 00:13:32,240 The returns diminish because at the beginning, we reduce the costs by a lot in just a few iterations. 208 00:13:32,810 --> 00:13:36,810 At the end we can do a lot of iterations, but we reduce the cost by almost nothing. 209 00:13:40,010 --> 00:13:45,950 An alternative is you could use a separate validation set and stop when the validation cost increases. 210 00:13:46,070 --> 00:13:48,110 But that is not the focus of this lecture. 211 00:13:48,800 --> 00:13:54,200 Unlike the training cost, the validation cost is not guaranteed to decrease at every round because 212 00:13:54,200 --> 00:13:55,660 that is not what we are minimizing. 213 00:13:56,000 --> 00:13:56,960 With respect to. 214 00:14:00,810 --> 00:14:06,450 You can also use the same plot to help improve your learning rate if the cost explodes or becomes not 215 00:14:06,450 --> 00:14:09,210 a number, your learning rate is probably too high. 216 00:14:09,840 --> 00:14:13,620 If your cause converges too slowly, you can try increasing your learning rate. 217 00:14:17,990 --> 00:14:19,940 So what does your final code look like? 218 00:14:21,870 --> 00:14:25,710 Well, we know we want to create a class with at least that predict and fit functions. 219 00:14:26,220 --> 00:14:30,870 You don't have to use a class, but I find that it encapsulates the code nicely and provides useful 220 00:14:30,870 --> 00:14:31,470 structure. 221 00:14:33,210 --> 00:14:38,280 Notice that once you have the algorithm in math, there isn't much work to convert it into no code. 222 00:14:38,970 --> 00:14:40,950 You know how to do matrix multiplication. 223 00:14:41,310 --> 00:14:42,210 You know how to add. 224 00:14:42,420 --> 00:14:43,530 You know how to subtract. 225 00:14:44,040 --> 00:14:47,370 These are basic arithmetic operations that I'm sure you know already. 226 00:14:48,240 --> 00:14:52,620 If you don't know how to do these and num pi, then I have a completely free number, of course, on 227 00:14:52,620 --> 00:14:54,300 the number I stack that you can take. 228 00:14:58,010 --> 00:15:02,690 Here is the second part of the code where we actually create an instance of the class and then use it. 229 00:15:03,410 --> 00:15:07,100 I've also included the squared error cost function just for completion sake. 230 00:15:12,320 --> 00:15:16,970 What's cool about this template is that it doesn't really change from one algorithm to the next. 231 00:15:17,480 --> 00:15:22,670 The structure is basically always going to be this way, at least for supervised learning and unsupervised 232 00:15:22,670 --> 00:15:23,030 learning. 233 00:15:23,930 --> 00:15:28,310 So as we discussed before, it is a matter what algorithm you're implementing, whether it be linear 234 00:15:28,310 --> 00:15:31,700 regression, logistic regression, a neural network and so on. 235 00:15:32,000 --> 00:15:37,550 They always have the predicts function and the fit function and learning what goes inside these functions 236 00:15:37,850 --> 00:15:39,880 is equivalent to learning the algorithm. 237 00:15:41,510 --> 00:15:46,910 So, in fact, one way to begin coding is to start with just this boilerplate code and then fill in 238 00:15:46,910 --> 00:15:47,600 the blanks.