1 00:00:02,170 --> 00:00:04,420 Everyone and welcome back to this class. 2 00:00:06,390 --> 00:00:10,620 This lecture is about how to write code independently when you're doing machine learning. 3 00:00:11,850 --> 00:00:16,379 You always hear me say that the first thing you should try to do when you learn an algorithm is to go 4 00:00:16,379 --> 00:00:20,130 and write it up in code by yourself before looking at someone else's code. 5 00:00:21,000 --> 00:00:25,050 For some students who are new to machine learning, it's not quite clear how to do that. 6 00:00:25,710 --> 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 --> 00:00:33,090 more people to try to code by themselves. 8 00:00:36,570 --> 00:00:41,340 I like to encapsulate this in the phrase when it's time to code, you must code. 9 00:00:42,030 --> 00:00:46,800 So remember this whenever you see some code that demonstrates the concepts we learned in this course, 10 00:00:47,430 --> 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 00:00:58,500 --> 00:01:02,910 Sometimes people act like they are going to learn tennis by reading a book about tennis. 13 00:01:03,570 --> 00:01:05,790 This is, of course, obviously not possible. 14 00:01:06,690 --> 00:01:11,280 You need to learn some things about tennis, which are pure fact and resigning your brain as conscious 15 00:01:11,280 --> 00:01:11,820 facts. 16 00:01:12,150 --> 00:01:15,510 But then you have to go out and practise those concepts on the tennis court. 17 00:01:16,290 --> 00:01:21,420 Once you've exhausted and mastered everything you know so far, then you can learn new techniques. 18 00:01:21,990 --> 00:01:26,240 So you have this continuous cycle of learning new techniques and then practicing them. 19 00:01:29,640 --> 00:01:35,370 Why is this because learning the technique as pure fact doesn't mean you truly understand it? 20 00:01:36,030 --> 00:01:40,020 Practicing the technique helps you think about the technique from a different perspective. 21 00:01:40,590 --> 00:01:46,170 Practicing again and again leads to new understanding to learning in the subconscious, which is muscle 22 00:01:46,170 --> 00:01:46,590 memory. 23 00:01:47,100 --> 00:01:48,570 You must go through this cycle. 24 00:01:49,440 --> 00:01:54,330 But some students naively think they are going to read an entire book about tennis and then be a tennis 25 00:01:54,330 --> 00:01:56,700 master the first time they try playing. 26 00:01:57,330 --> 00:02:02,640 This is a very common, especially when all you really need to do is sit there and watch a video and 27 00:02:02,640 --> 00:02:03,960 no one can force you to type. 28 00:02:04,740 --> 00:02:06,810 A lot of people don't even think to try. 29 00:02:07,770 --> 00:02:13,080 So just remember this whenever you see code, when it's time to code, you must code. 30 00:02:16,640 --> 00:02:20,270 First, we are going to talk a little bit about why you want to code by yourself. 31 00:02:21,170 --> 00:02:26,270 A lot of the time what an algorithm does or how it works isn't quite clear the first time you learn 32 00:02:26,270 --> 00:02:26,630 about it. 33 00:02:27,440 --> 00:02:32,450 Every individual will have their own unique background, so they might be familiar with the patterns 34 00:02:32,450 --> 00:02:33,740 and know exactly what to do. 35 00:02:34,130 --> 00:02:37,100 While some individuals might have questions that no one else has. 36 00:02:37,940 --> 00:02:43,190 Sometimes you gloss over details because you assume you know what's going on when really there are important 37 00:02:43,190 --> 00:02:44,480 things you're not considering. 38 00:02:45,020 --> 00:02:47,930 So this is how trying to code by yourself can help you. 39 00:02:48,770 --> 00:02:52,100 Coding by yourself forces you to think about each and every detail. 40 00:02:52,520 --> 00:02:54,440 It forces you to think line by line. 41 00:02:55,820 --> 00:03:00,500 You have to be familiar with the data types and the shapes of all your variables, and they all have 42 00:03:00,500 --> 00:03:01,790 to fit together properly. 43 00:03:02,120 --> 00:03:03,530 Kind of like Lego blocks. 44 00:03:07,380 --> 00:03:13,110 As an example, you know that in order to do element wise matrix additions, both matrices have to have 45 00:03:13,110 --> 00:03:14,310 the exact same shape. 46 00:03:15,060 --> 00:03:19,980 So if you find that they are not the same shape, then one of your previous assumptions was incorrect. 47 00:03:21,000 --> 00:03:22,710 So you should go back and correct it. 48 00:03:23,760 --> 00:03:28,260 Lego blocks have to follow a specific set of rules in order to fit together properly. 49 00:03:29,040 --> 00:03:34,110 If you make incorrect assumptions about how Lego blocks fit together and you try to join them, it's 50 00:03:34,110 --> 00:03:37,800 not going to work without trying to build things by yourself. 51 00:03:38,160 --> 00:03:40,020 You will never discover these details. 52 00:03:43,850 --> 00:03:46,280 Let's now talk about how to code by yourself. 53 00:03:47,000 --> 00:03:49,520 Consider the supervised machine learning scenario. 54 00:03:50,210 --> 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 55 00:03:55,370 --> 00:03:57,980 from the inputs that are very close to the targets. 56 00:03:58,730 --> 00:04:01,760 We typically call the inputs X and the targets Y. 57 00:04:02,450 --> 00:04:06,980 Sometimes we call the targets T, but for the purpose of this lecture, we'll call the targets Y. 58 00:04:09,950 --> 00:04:11,750 Now, this next point is a key point. 59 00:04:12,440 --> 00:04:16,070 It doesn't actually matter what sex is and what, why is it? 60 00:04:16,070 --> 00:04:16,670 Does it matter? 61 00:04:16,670 --> 00:04:22,160 If you're looking at an e-commerce data set and the columns of X might be time on site, time of day? 62 00:04:22,400 --> 00:04:24,620 How many pages the user looked at and so on. 63 00:04:25,220 --> 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 64 00:04:30,740 --> 00:04:31,340 of an image. 65 00:04:32,000 --> 00:04:33,140 This is the key point. 66 00:04:33,470 --> 00:04:37,620 We say all data is the same when we do linear regression. 67 00:04:37,640 --> 00:04:41,630 We don't have different kinds of linear regression for e-commerce and X-ray images. 68 00:04:42,170 --> 00:04:45,920 Linear regression is the same algorithm, no matter what your data set is. 69 00:04:46,340 --> 00:04:48,290 So we say all data is the same. 70 00:04:49,370 --> 00:04:54,800 Theoretically, I could give you a data set of X's and whys, and you could train a classification algorithm 71 00:04:54,800 --> 00:04:58,460 on it without me even telling you what the X's and Y's mean. 72 00:04:59,120 --> 00:05:01,190 You should get very comfortable with this idea. 73 00:05:04,520 --> 00:05:10,010 Sometimes people say, well, this isn't practical because I want to do concrete examples, but that's 74 00:05:10,010 --> 00:05:11,960 because they're not thinking intelligently. 75 00:05:12,560 --> 00:05:17,390 In fact, this is the most practical thing we can do because it means everything we learn. 76 00:05:17,630 --> 00:05:20,180 We can apply it to any data set that exists. 77 00:05:20,870 --> 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 78 00:05:26,270 --> 00:05:29,330 advertising dataset without learning anything new at all. 79 00:05:29,900 --> 00:05:32,330 It's truly the greatest form of lazy programming. 80 00:05:32,750 --> 00:05:35,480 Learn something once and apply it to every industry. 81 00:05:39,070 --> 00:05:45,070 One great consequence of all data is the same is that there is an unlimited amount of practice opportunity 82 00:05:45,070 --> 00:05:45,460 for you. 83 00:05:46,270 --> 00:05:51,670 You can download data sets from Kaggle, from Google, from Wikipedia, from Amazon or from anywhere 84 00:05:51,670 --> 00:05:53,830 else and try the algorithms you learned. 85 00:05:54,640 --> 00:05:57,010 What this means is, let's say you have a data set. 86 00:05:57,280 --> 00:06:00,580 You care about more than something we do in Class C. 87 00:06:00,580 --> 00:06:03,640 In class, we have to use data sets that everybody can understand. 88 00:06:04,180 --> 00:06:08,440 For example, text and images, everybody should know what text and images are. 89 00:06:09,220 --> 00:06:12,430 Text and images are just fundamental data types on the web. 90 00:06:13,000 --> 00:06:17,830 You probably don't even think about the fact that text and images are data because it's so trivial. 91 00:06:18,820 --> 00:06:23,080 In any case, we work with text and images a lot because everybody understands them. 92 00:06:23,860 --> 00:06:29,710 But let's say you're a biologist and you know about DNA, you find DNA and genomics really interesting. 93 00:06:30,370 --> 00:06:32,590 So you want to use the algorithms we learn on that. 94 00:06:33,190 --> 00:06:33,880 Well, that's great. 95 00:06:33,880 --> 00:06:34,990 And you should totally do it. 96 00:06:35,680 --> 00:06:41,410 Remember, all data is the same, so all you need to do is convert your data into the appropriate XS 97 00:06:41,410 --> 00:06:44,410 in ways to plug in to our machine learning algorithms. 98 00:06:45,700 --> 00:06:52,060 The downside is we most likely won't talk about DNA and class unless it's a very simple example, because 99 00:06:52,060 --> 00:06:57,240 most computer scientists are not biologists too, so they don't know what DNA is. 100 00:06:57,280 --> 00:06:58,960 They don't understand the specifics. 101 00:06:59,470 --> 00:07:01,570 So an example wouldn't make much sense to them. 102 00:07:02,320 --> 00:07:05,530 This is opposed to text and images, which makes sense to everybody. 103 00:07:06,370 --> 00:07:12,070 Same goes with any other specialized field like finance, computer networking, cosmology and so on. 104 00:07:16,870 --> 00:07:20,410 So now we can start up from the right place to talk about supervised learning. 105 00:07:20,830 --> 00:07:22,750 We have our exes and we have our wives. 106 00:07:22,840 --> 00:07:26,040 What do we do with them in supervised learning? 107 00:07:26,050 --> 00:07:28,300 We know that there are two main things we want to do. 108 00:07:28,960 --> 00:07:33,070 We want to do training and we want to do prediction and psyche to learn. 109 00:07:33,070 --> 00:07:34,900 All the models have the same API. 110 00:07:35,350 --> 00:07:39,820 It doesn't matter if you're doing logistic regression or decision trees or random forest. 111 00:07:40,270 --> 00:07:45,220 All supervised models in psychic learning have the same two functions fit and predict. 112 00:07:45,880 --> 00:07:48,550 Fitting a model is just a synonym for training a model. 113 00:07:49,090 --> 00:07:53,020 If you don't believe me, you can go look at this Typekit documentation yourself. 114 00:07:56,570 --> 00:08:01,670 When you're learning about supervised machine learning, what you're really learning is what goes inside 115 00:08:01,670 --> 00:08:02,630 these two functions. 116 00:08:03,110 --> 00:08:05,840 What are the parameters and how are those parameters learned? 117 00:08:06,590 --> 00:08:10,790 That's really what all supervised machine learning is filling in these two functions. 118 00:08:11,480 --> 00:08:16,280 Taking this perspective is going to give you a code structure, and it's going to make things much easier 119 00:08:16,550 --> 00:08:17,870 to visualize in your head. 120 00:08:20,950 --> 00:08:24,100 Let's now do a simple example to solidify this idea. 121 00:08:24,970 --> 00:08:29,290 And this example, I'm going to give you an algorithm and you will implement it in code. 122 00:08:30,040 --> 00:08:33,280 There are two key points I want to make before I give you the algorithm. 123 00:08:34,000 --> 00:08:37,690 First, I'm not going to give you much intuition about how it works. 124 00:08:38,320 --> 00:08:41,710 Second, I'm not going to derive the theory behind why it works. 125 00:08:42,460 --> 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 126 00:08:47,620 --> 00:08:51,610 pieces of information in order to translate pseudo code into code. 127 00:08:52,330 --> 00:08:58,870 A lot of time, these three concepts intuition, theory and implementation work to reinforce each other. 128 00:08:59,680 --> 00:09:05,270 The reason I'm focusing on implementation right now is because it's the one people most often miss or 129 00:09:05,290 --> 00:09:10,180 think they don't need at all when really it's an extremely important part of the learning process. 130 00:09:13,610 --> 00:09:18,770 OK, so the pseudocode is as follows in my predict function, as you already know. 131 00:09:18,800 --> 00:09:24,830 I'm going to take in some input data x my predictions will be y hat equals x times w. 132 00:09:25,700 --> 00:09:28,400 You might already recognize this as linear regression. 133 00:09:29,090 --> 00:09:34,220 However, forget what you know about linear regression and pretend that this is just some formula I 134 00:09:34,220 --> 00:09:34,700 gave you. 135 00:09:35,510 --> 00:09:38,960 What I can tell you is that we are creating some kind of regression model. 136 00:09:39,710 --> 00:09:44,210 It should be clear from this equation that the parameters of the model are contained in this vector 137 00:09:44,210 --> 00:09:44,630 of weights. 138 00:09:44,630 --> 00:09:50,990 W I would probably also tell you that W is a vector that's the same size as the number of feature vectors 139 00:09:50,990 --> 00:09:51,560 in X. 140 00:09:52,250 --> 00:09:56,000 Of course, this must be true in order for the matrix multiplied to be valid. 141 00:09:56,870 --> 00:10:01,100 This acts is kind of a sanity check, so you can make sure everything is making sense. 142 00:10:04,460 --> 00:10:10,700 In my fit function, I want to perform this loop some number of times, since we know from the previous 143 00:10:10,700 --> 00:10:15,710 slide that the parameters of the model are contained and W, then it should be clear that inside the 144 00:10:15,710 --> 00:10:18,380 fit function, what we want to do is update W. 145 00:10:19,100 --> 00:10:21,440 Not surprisingly, that's what's happening here. 146 00:10:21,710 --> 00:10:27,090 Another good sanity check now because you're already familiar with linear regression. 147 00:10:27,140 --> 00:10:30,110 You probably already recognize this as gradient descent. 148 00:10:30,740 --> 00:10:34,700 Again, I want you to pretend that you don't know that, but what do you know? 149 00:10:38,540 --> 00:10:43,490 You know, that iterative algorithms are common in machine learning and that gradient descent is among 150 00:10:43,490 --> 00:10:44,330 the most common. 151 00:10:44,930 --> 00:10:49,790 You could probably infer that this is some form of gradient descent, but you don't need to know what 152 00:10:49,790 --> 00:10:51,560 it is in order to write it in code. 153 00:10:52,310 --> 00:10:53,270 What else do we know? 154 00:10:54,170 --> 00:10:59,300 You know what a for loop is and you know how to write when in Python, you know that X is an end by 155 00:10:59,300 --> 00:11:04,520 the matrix containing the input data, and that Y is an inside vector containing the targets. 156 00:11:05,480 --> 00:11:11,270 You know, that white hat is an enzyme vector containing the predictions, you know, that W is a disease 157 00:11:11,270 --> 00:11:13,220 vector containing the model parameters. 158 00:11:13,730 --> 00:11:19,070 And you know how to add, subtract and multiply these objects so we can write this algorithm without 159 00:11:19,070 --> 00:11:20,420 even knowing how it works. 160 00:11:21,320 --> 00:11:24,680 Of course, a lot about how it works can just be inferred from this formula. 161 00:11:28,790 --> 00:11:30,050 Here is what you don't know. 162 00:11:30,410 --> 00:11:33,410 You don't know the number of times the loop is supposed to run. 163 00:11:33,890 --> 00:11:35,540 You don't know Ada the learning rate. 164 00:11:36,110 --> 00:11:37,430 This is completely fine. 165 00:11:38,000 --> 00:11:40,970 You can't let this lack of knowledge stop you in your tracks. 166 00:11:41,750 --> 00:11:46,550 The truth of the matter is they're are going to be situations where you're not told the exact numbers 167 00:11:46,550 --> 00:11:47,210 to plug in. 168 00:11:47,810 --> 00:11:51,530 In many cases, the answer is going to be it depends on the problem. 169 00:11:52,400 --> 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 170 00:11:58,460 --> 00:12:00,560 guesses based on what you already know. 171 00:12:04,130 --> 00:12:08,900 So what can you do when you know that the learning rate is supposed to be a small number? 172 00:12:09,260 --> 00:12:09,950 How small? 173 00:12:10,160 --> 00:12:11,510 It depends on the problem. 174 00:12:12,290 --> 00:12:15,290 Typically, it's a number less than one, like zero point one. 175 00:12:15,860 --> 00:12:17,900 If that doesn't work, you might try lowering it. 176 00:12:18,500 --> 00:12:20,660 Typically, we lower it by factors of 10. 177 00:12:21,290 --> 00:12:26,240 So, for example, the next learning rate we would try is 10 to the minus to 10 to the minus three and 178 00:12:26,240 --> 00:12:26,780 so on. 179 00:12:30,930 --> 00:12:36,330 Typically in machine learning, whether you're looking at a supervised or unsupervised algorithm, there 180 00:12:36,330 --> 00:12:39,930 is an associated cost or objective function that you're trying to minimize. 181 00:12:40,770 --> 00:12:45,270 Let's suppose I tell you it's squared error, which is typically what we would use for regression. 182 00:12:46,020 --> 00:12:50,370 Well, now you have a way to choose the number of iterations of the loop and the learning rate. 183 00:12:50,910 --> 00:12:51,840 How do we do this? 184 00:12:56,190 --> 00:13:00,570 So inside your fit function, you would plot the cost as a function of iteration. 185 00:13:01,320 --> 00:13:04,650 I always recommend doing this no matter what algorithm you're looking at. 186 00:13:05,550 --> 00:13:08,070 In general, you always want the cost to converge. 187 00:13:08,790 --> 00:13:13,980 The pattern you typically see is that there is a steep drop at the beginning, and it slowly flattens 188 00:13:13,980 --> 00:13:19,740 out as the number of iterations increases to choose the number of iterations you want to stop when the 189 00:13:19,740 --> 00:13:21,330 curve is sufficiently flat. 190 00:13:22,080 --> 00:13:25,320 You might recognize this as a diminishing returns scenario. 191 00:13:26,420 --> 00:13:32,180 The returns diminish because at the beginning, we reduce the cost by a lot in just a few iterations. 192 00:13:32,720 --> 00:13:36,800 At the end, we can do a lot of iterations, but we reduce the cost by almost nothing. 193 00:13:39,920 --> 00:13:45,920 An alternative is you could use a separate validation set and stop when the validation cost increases. 194 00:13:46,040 --> 00:13:48,110 But that is not the focus of this lecture. 195 00:13:48,740 --> 00:13:54,200 Unlike the training cost, the validation cost is not guaranteed to decrease at every round because 196 00:13:54,200 --> 00:13:56,900 that is not what we are minimizing with respect to. 197 00:14:00,720 --> 00:14:06,450 You can also use the same plot to help improve your learning rate if the cost explodes or becomes not 198 00:14:06,450 --> 00:14:06,990 a no. 199 00:14:07,230 --> 00:14:11,670 Your lending rate is probably too high if your cost converges too slowly. 200 00:14:11,700 --> 00:14:13,620 You can try increasing your learning rate. 201 00:14:17,930 --> 00:14:19,910 So what does your final code look like? 202 00:14:21,810 --> 00:14:25,680 Well, we know we want to create a class with at least that predict and fit functions. 203 00:14:26,160 --> 00:14:30,870 You don't have to use a class, but I find that it encapsulates the code nicely and provides useful 204 00:14:30,870 --> 00:14:31,440 structure. 205 00:14:33,120 --> 00:14:38,250 Notice that once you have the algorithm in math, there isn't much work to convert it into numpty code, 206 00:14:38,880 --> 00:14:43,470 you know how to do matrix multiplication, you know how to add, you know, how to subtract. 207 00:14:43,980 --> 00:14:47,340 These are basic arithmetic operations that I'm sure you know already. 208 00:14:48,150 --> 00:14:53,100 If you don't know how to do this in Nampai, then I have a completely free Nampai course on the Nampai 209 00:14:53,100 --> 00:14:54,270 stack that you can take. 210 00:14:57,980 --> 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. 211 00:15:03,350 --> 00:15:07,070 I've also included the squared error cost function just for completion sake. 212 00:15:12,200 --> 00:15:16,910 What's cool about this template is that it doesn't really change from one algorithm to the next. 213 00:15:17,420 --> 00:15:22,640 The structure is basically always going to be this way, at least for supervised learning and unsupervised 214 00:15:22,640 --> 00:15:23,030 learning. 215 00:15:23,870 --> 00:15:28,310 So as we discussed before, it doesn't matter what algorithm you're implementing, whether it be linear 216 00:15:28,310 --> 00:15:31,720 regression, logistic regression, a neural network and so on. 217 00:15:31,940 --> 00:15:37,550 They always have the predict function and the fit function, and learning what goes inside these functions 218 00:15:37,790 --> 00:15:39,860 is equivalent to learning the algorithm. 219 00:15:41,450 --> 00:15:47,000 So in fact, one way to begin cutting is to start with just this boilerplate code and then fill in the 220 00:15:47,000 --> 00:15:47,540 blanks.