A Short Introduction to Machine Learning | Daniel M. | Skillshare

A Short Introduction to Machine Learning

Daniel M.

Play Speed
  • 0.5x
  • 1x (Normal)
  • 1.25x
  • 1.5x
  • 2x
20 Lessons (1h 11m)
    • 1. Introduction

      1:34
    • 2. Defining Machine Learning

      3:35
    • 3. Understanding Machine Learning (Part 1)

      3:41
    • 4. Understanding Machine Learning (Part 2)

      3:38
    • 5. Understanding Machine Learning (Part 3)

      3:28
    • 6. Core Concepts (Part 1)

      3:29
    • 7. Core Concepts (Part 2)

      3:52
    • 8. Core Concepts (Part 3)

      3:58
    • 9. Core Concepts (Part 4)

      3:49
    • 10. Algorithms: Decision Trees

      4:04
    • 11. Algorithms: K Means Clustering

      4:02
    • 12. Algorithms: K Nearest Neighbor

      4:05
    • 13. Algorithms: Naive Bayes

      4:07
    • 14. Algorithms: Regression

      4:06
    • 15. Problems That Utilize Machine Learning

      3:35
    • 16. Choosing the Right Algorithm

      3:48
    • 17. Fitting the Data

      3:31
    • 18. Following the Data

      4:04
    • 19. Usual Challenges

      4:04
    • 20. Next Steps

      0:44
15 students are watching this class

About This Class

Machine learning is one of the most important areas of Artificial Intelligence. It provides developing methods that can automatically detect patterns in data and then use the uncovered patterns to predict future data.  It can be applied across many industries to increase profits, reduce costs, and improve customer experiences.

In this class I’m going to provide you with a short introduction to the field of machine learning.It will make the fundamentals and algorithms of machine learning accessible to anyone. This means plain-English explanations and no coding experience required.

Transcripts

1. Introduction: machine learning is one of the most active areas, not deficient diligence. That's partly due to the explosion of big data. But it's also because of the huge advances notionally our workers. We now have motions that I love to drive cars, search for Parma suit, person, even be expert players in games that require complex strategy and creativity. Disabilities are just the beginning. My issues are getting better, a task that in the past would only be accomplished by humors machine learning how to make better judgment pattern matching its strategic decision making. This course will give you a high level overview of machine learning Carstensen technology course to see how much is learnt. Did you see some of the most part? They use machine learning our weapons. Finally, you get someone isn't hopeless. Take the best. Many organizations are taking advantage of this technology and say that one of the biggest obstacles is finding employees that intangible about machine learning. This isn't just for data science is, but also for managers and even executives who could benefit from nine more about this technology. That's why it's courses designed for managers on drip in your students or professionals who want to better understand machine learning. You see how much learning fits into the overall world of artificial intelligence? Did you see how to harness this in technology to make better decisions and find patterns in the data? So let's step into the world of machine learning. 2. Defining Machine Learning: machine learning has been around for a long time. You can probably tell because it turned itself is retold. Passion You don't after your computers and purples machines. The Time machine learning got its start in 1959 or computer partner, our two. Samuel wanted live computer school and the behavior. Instead of being programs to do specific tasks, this type of thinking was drastically different from almost computer scientists. COMPUTERS. A computer needs to be called exactly what Go think about how you there. Apple computers. Most programs are a series of explicit instructions. That's why we don't pretty soccer for something like a baking application. You be very size, my creative destruction that says something like you for customer straight on the money and that exceed their balance, they cancelled the transaction. That's an explicit instruction. If you see Extender, why machine learning is there. Here you're not creating detailed instructions is that you're giving the computer the daytime dozing needs to study the problem and solve it. Stopping for what a ball. Then you're giving the computer the ability to remember what it did. Sir can adopt. People learn. That's not that much defense. Hawking was lower. Several years ago, I decided to purchase a new bookshop with my wife and would be couple very large off my Kia something up on the box. I realize that putting it together would be incredibly difficult. I needed the have to learn by doing it, and I make a lot of these days. The instructions say that each one of the show's has something called a dollar on each side , so I put hours into each of the falls. Then I shot the shock into the sides and with people my wife put under the show, and we're post processing that it were. Once I figured out the first show, I had enough experience to start the other shelves. That way I didn't have to look at the instructions Each time I start a new show, this human learning was about starting small than going big. I hit the problem, so I formed the basis of my own experience. I didn't use that. Want to try some larger passion? Can I waited for feedback to see departmental Justin alone. If it doesn't piers this time, other than you know that work in other to my man, Were you Once the paste, I could probably put together many different types of like your furniture machine learning computer slow bodies in much the same way the machine starts by tasting something small of excess moral part of the baby. Then it uses a statistical arboretum to see how the date up together as a human hater with at my shop in the dollar. But machine haven't are within that say that that sort data should be treated the same way the machine would then use it on with look for patterns. The machine will then get some feedback. Remember that we tested the show by pushing it down to see equal supportable here. The motion, my test that I'll come against the training data to see if it was going there. Everything the mission or something you get past the database, in a sense, is starting it in his long term memory circular. Looking at that, keep in mind that the machine the human both came out with other expert is I learned how to put together a needlessly complex bookshop. So now I know much more about furniture assembly, and I machine with much more about data 3. Understanding Machine Learning (Part 1): you may think abortion learning as a different impulse, something that's already around. Maybe it's just an updated way, describes it tasted or anywhere to talk about it a science, but that you didn't think about Machine learning is the focus of the term learning machine . Learning certainly has to face this. It might also be a key part on your little science papers, but these are just two of emotion is to learn they're not a substitute for learning. Think about what it means to learn. What are the different strategies that you used to know something you know? How can it take these strategies? And in a pleasant to machines, imagine you wanted to learn how to play chairs. You could do this a couple of different ways. You could, Sarah, just you. They would introduce you to some of the different chest pieces and how they move across the poor. We could practice, but great ingested people, and they were supervising moves and how you would have made him stay. Then after a while, you do that with finished all your lessons, and it's time to play competitive others. Let's say I couldn't find you there then it would go to public parks. And what several 100 expert players in the game you couldn't ask them? Questions just quietly Want to learn. If you do this long enough, you probably understand the game. You might not know the names of the chess pieces, but you could understand the most in strategies from your powers of observation. We might even try a combination of these two approaches. Adjust Daughter will show the basic rules. Then you go back and watch other people playing. You'd have a high level overview and the names of the chess pieces, but your land observations as a way to see the strategies in improved this tree strategies are very single hall motion with you could have something called supervised learning. Here, a data scientist acts like a total for the machine. The gentle machine. By showing get basic rules and giving it an overall strategy. You could also try unsupervised learning. You just said that we should make all the observations on its own. The machine might not more of the different names and labels, but there are matters in their own. Finally, you could make the toe and try semi supervised learning you're change the machine just a little bit, so they get a valuable overview. The most of the learning about the rules and strategies is drops every living patterns. As you can imagine, all three of cultures have their own strengths and weaknesses for supervised learning. Getting the logical Cooper there needs to be someone other than knows a lot about chasing crucial hope. To play the game with unsupervised learning, you need to have access lots of data. You might not be able to go toe part. Watch hundreds of different experts that game. It also depends, right Orbital. We watch. You need to watch people partly well, who sent me supervised learning. You give him trouble on both sides. If you have a bad photo, it would be much more difficult. Challenge observations, on the other hand, if you have a quick there, but the people who have self are poor players. They don't understand the game but not really become master players. My people position working, decide which approach works fast, but after we just have to do the best with what's available. If you can find that photo, they have to do your best. Bottom seven people in public parks. If you don't have a public park, they have to do your best for the good. You can only go semi supervised learning if you have access to bowl. 4. Understanding Machine Learning (Part 2): machine learning is already used to many different industries. Want to check the weather? Or if you're typing something into the search engine, you don't know dependent. Getting commission everything. An organization that has a lot of data and is looking for better ways. Understanding can benefit from this technology, but there are some areas that they're much the book here. Think about this Web application right now. You're either watching this video when your mobile device on the computer machine learning can help us to help you put your experience. This application can put a huge amounts of paper. When you click on a different veto, it can equip about watching. When you are in the time of the day, it's pretty easy to correct massive amounts of paper. What's difficult is getting insights from the paper machine. Learning can take all the data about you. You could CBD is that they're much more relevant what you like. Maybe in the search for something new who could have a custom? I search based on your dangers. This application can start. Your behavior is a fuckin better meter years. It might seem strange, but computers getting annoying better has actually turned up some of the most lucrative businesses. Companies like Google pays apple and linking aren't using machine learning to better understand you. Some companies are people pushing the envelope of privacy by looking for your previous searches. Are trying to better understand your friends that way? Every time you use their service, they can give you a customized experience. That's why your Facebook newsfeed isn't it to you. It's also I took people putting in the exact same search time to cook and my get completely different results. Anything you know, website is something that says Recommended for you. Then you're probably benefiting for machine learning. Amazon toe other previous purchases and then uses machine learning are becoming tired limbs . Netflix uses machine learning the complex path or circle recommended shows based on what already watch you to bless up the next video. Based on those reports in the past, the business as a way to make better recommendations for you. Many organizations are using machine learning for automatic translation. You took my was natural language processing trust capital beating in except those their trust. I think it's pitch into bags, some said, with his machine learning to translate once it off subtitles into several different languages in machine learning. Greasy artificial intelligence to help the program, five batters and massive data sets after the camp patters in data that you must never seen . That's one of the most interesting things about machine learning. He's not just a chance form of human learning, they said. It's a completely different way to find patterns, make decisions and getting rid of insights. So if you want to use machine learning in your organization, you have to think about the way the machine laws. That way, you can stop the correct date that will live a program to better understand your past more . Before you start the machine learning project, you have to think about your data. Is it a high quality? Do you have enough data to know something you remember that this is the way of machine learning program will be the world, the brother of the view, the more likely it is to find something interesting. You don't want your program looking for key for when land use machine learning, start thinking about the best parties to get high party in the first data stairs 5. Understanding Machine Learning (Part 3): much of computer science is still about working with this. Places. Structures in traditional programming computer emotion, talk sector input, produce allow person with the input is the commander, and the output is a predetermined, its parts that will work well when you have a program for populations. But it gets a little trickier when human can explicitly instruct the computer wanted bull. In these cases, you the programming mother that they lost, the motion learned. You also have to give the machine some ability to respond to beat back. She's a perfect scenario for machine learning. Imagine it, creating a program that Mr Takes Pamela's juice. These messages are usually filled with I wanted advertisement of people viruses. You can easily create a world with the program that lets messages with common spam words they can kill. The world's like called a lot of your winner. This was certainly a lot of spam messages, but he will be pretty easy to school. You can change the world lottery to contain zero or just use images. It also might lead a lot of post positives. Maybe your friends and your joke about putting a lot to be this inner would actually have to get really good. This step off challenges don't work well when were limited to carefully depart instructions . You can't simply create an input command with a pretty Temer's pars. That's what machine learning switches things around. Instead of importing instructions, reliable data he's sitting pretty virus past. You'd be working with machine learning are with him to help the motion that help respond. To start, you want to speak the data into the state and training data. The training data is a small chance that you find patterns. Sometimes a mother will have much should make sense of the day that was a physical are this are worth and suddenly should make accurate predictions or see patterns between different parts of your data. Let's think about how much you learning my work with this bomb program. Let's take 10,000 email messages, Search Engagement said. Well, listen to build the refined our mother before we try to not a state of over a 1,000,000 messages, you can use your face. Take that should the motion. Different examples of palm The mucus is a classic for emotion. Living apart wouldn't have hesitated. Email little groups. You could have a stamp in the regular message. This is often corporate classification. The machine is by finding comes off force that are more likely to be calm. Respond messages. Then it comes up with the score to show the likelihood it bam As a machine learning expert , you need to decide the best once acquired Albert in yours, then with the hyper parameters of our Britain. And the motion is a political job predicting whether or not that even messages. Pam, once you're satisfied and that will be your initial data model, use a machine learning problem and trendiness city hyper parameters to make an accurate prediction. The key thing to remember is that even though the programmer impose a date estimates that are going to make corrections, it's ultimately the machine that makes the decision about whether messages pound. In some cases, the programmer might not given a hollow machine. Learning was 6. Core Concepts (Part 1): well known up about a date. I have machine calling the doctor supervised learning. He supervised learning you should the machine the connection between different variables and no outcomes in machine learning. This is called the Table Table Data in the Great Power. It's called label data because it's over the top with identifying information. Imagine you wanted to train a shame to help you predict how long it will take you to drive home. You will start by quitting the set of labeled data his state. That would be good weather conditions, time of day and whether it's a hard day. These are being balls the art, but will be the amount of time to drive home on that particular day. In this case, they're independent. Variables will be the input and independent variables. Who did the out for here? Won't use different machine learning algorithms. Tomato relationship between these different variables, you can use statistical regression for this and determine how the independent variables affect the dependent variable. He didn't know the training outside. It would take longer to drive home, but machines have to rely on data and statistics. They have to look at the length of their commute and compared it to label data hasn't weather. Let's how can create a supervised motion learning model to help us determine welcome you. The first thing you want to do is to quit, the dream said. Based on this training said, your machine must see that there is a direct relationship between the model train in the time it takes to get home. The morning is, the longer will be on the road. It might also see there's a connection between family work and that will be on the road. The closer you are to 5 p.m. The longer it takes to get home your motion, find some of the relationship with labor data. This is the start of our data model. It begins to understand concepts like holding impacts the way people drive. It also starts to see the doctor. More people would travel during certain times of day. Then your motion takes his training said in a plastic focus it of your data. It looks set the mother in this training, said host. So when you look at many more days, you could ask the computer to predict how long you they could drive from each day and then give it feedback on how accurate it was. This prediction over time the machine learning adapt its mother people are your mission will continue to do on the mother in the train said, And you can make judgments based on the new data. This is pretty similar to Hawking was lower. Think about when you learn how to drive. He started in this more parking lot with you. Then he got to the point where we could have productively and safely in this more space. Once you put up enough confidence you about in the open floor, we want everything you need to know about driving in the park. Not, but you should get an update that help you get started. Chances are you still have a lot of years and hopefully you continue being pulled over time until you become an expert driver. The kitten to remember Is that a superb admission? Learning a lot more about the training data. You keep your table data into the machine that isn't classified, so don't commit home. You know much more about the weather holiday some time of day. This label data is the key difference between supervised learning in other parts of machine learning. 7. Core Concepts (Part 2): most supervised early mice in the most from you earlier if you think about it, weren't actually that many of the things I already know you weren't actually thought. Hope we'll have to have your hands is that you learn much of this to study in observation, Learning and improving my child error is that you don't supervised learning. Nice to provides There. You're not working with people Date that you're not showing the machine. The quick answer is that the reason different Albertans read the motion create connections but starting and observing the data. Think of it this way. Every Halloween the past few years myself Is that something called trick or treating? All the kids in the neighborhood go door to door and that's for candy. At the end, he comes home with a bag of conscious of the Began These the person he wants to do is classify the candy bowl. What, like spares in machine learning. This is called multi class specification. There are several different groups and would you want to pacify your day Now? In the past month, some Hispanic with for my experience, I've been able to supervise this learning I can help create the awful catbird's. But candy such as chocolate peanut butter means or Gannon's even showed him some of the tax and neighbors for his data, says means are almost always small in this one at a good pace. Peanut butter Candace didn't look like chocolate markets approach is almost always have a terrible orange packaging. Then I quit the small training set what I pull a few of his Candace into separately, the girl. Then we can classify some of the unknown Candies into different categories. After I leave the room, I can hear him saying me and chop me and I see classify, There's of the bag Now my son has been. Is that even a different country? They feel badly they can participate in this tradition. So each year this Emerson, a large pickup, local candy they collected or here with this bag you can't use a foremost provides learning because it's enable data. Neither he nor I have ever seen candy before, and the rappers are nothing but you can't understand. So in this case, we don't form of unsupervised learning. He looks at the bag, increase his own way of trusting the candy. He does this vessel. Usual study of the data and observations. It might be a group based on the size or the terror of the wrapping. In fact, one year he created a grouping that would never see them. It was a small plaster that he called from candy. It was canIs ever made out of power glasses. This is a key part of UN supervised him. Like my son. The machine studies the data in, comes up with these observations. In this case, it was about casting that candy into the schools. But you can also use this technique to create connections of people. Pass something new you might see. Jock companies used this technique to drive plant chemicals that might behave like an antibiotic. They could use supervised learning. They should. The machine, a small training, said where the labour certain came because antibiotics. But they also must send a better chance of finding something new. If the Chapmans provides turning, they could have the machine study and observed millions of different chemicals. Then they could have the machine. Creator of clusters, maybe didn't find that class the with, you know, antibiotics. As you can imagine, one of the key things with unsupervised learning is access to massive amounts of data. The more data you have, the facilities for the machine top servants that change that money work faster. 8. Core Concepts (Part 3): semi supervised learning is across over the takes advantage of all supervising and supervised learning. Imagine you wanted to quit the program that would translate into text. It will be difficult to use all supervised learning. Your emotional plaster wants together. But wouldn't they say You know what this was? Me. Maybe, you know, she could identify each time someone said the word Koldo. The problem is this cluster in and classify the program doesn't know that this sample photo is the same as the world photo on the flipside you wouldn't want to only use provides learning. It would be part of labor intensive to label all of data. But what's in there sounds. Plus we could never created, Jennings said. That was not cannot be trusted all the possible words and phrases. So is that you much I semi supervised learning. You start with a small training set or very common Wharton phrases. The motion quit uses the performance of misclassification. This will be left training, said unsupervised learning. But then you were training set into the machine and allow it to study in surgeries of the data. This is similar to a supervised learning that will not a motion to expand its vocabulary based on most of the class five with the label data the program, which is a form of inductive reasoning to try and expand its over cavalry. It might use the training data that peninsula for Toto the Southport, or they come up with petrol stations. Maybe it can't use the world photograph, talk, photocopy or even important pedo. It would have the training data for the World Photo in the Sun photo, and so the motion with just using that increasing. We're making the translations. Sometimes she interest is in your connection. It does its best to trying with the model. So if the train station is correct, it will battle of these words. You usually get speed back for one observers. That's why when you get the translation, you laugh. Concealing that says how this translation, but in that learning is not only for more some unsupervised learning. There's another formal, semi supervised learning for just at the reasoning this plaza to narrow down the honorable data about thinking about what you already know about the collection. So if you're using transacted reasoning to just kinda post mail, you have to think about the data in a larger context. Think about some of the things that people say in a voice mail message. People probably put many knows that their voicemail messages they're giving information about people place the paying. They won't let me to talk about times and days. A difficult voicemail message myself. Something like this. Hi, this is Joe. I was going to come from a meeting next Tuesday. Give me a quote. There's a change. This is much deeper from a random sampling off audio. Then you might get from listening to people in a restaurant so transacted reasoning, trustee of the model, by making better cases about what in the unlabeled data, semi supervised learning is not that common, but it works for in certain areas it afterwards, better when it's not good for the machine to create so called. But at the same time, the data is so large that it's not practical to use supervised learning. We might see some unsupervised learning by clasping where pages were grouping together pictures or something like insects. The key thing to remember is that semi super best there are really only makes sense when the other two approaches have difficult serving a challenge. You might also keep in mind that interactive transacted prison money greater errors in the stable data. So try to think of them as the best solution for a particular problem and not, they say, the best papers talk with their machine learning challenges. 9. Core Concepts (Part 4): one type of machine learning has gotten a lot of attention over the last few years. It's called reinforcement. Learning. Reinforcement. Learning is different from supervised, unsupervised and supervised learning. With each of these techniques, you're trying to invent the best model. I want to buy the mother that can accurately classify different data sets for five minutes plasters. Once you have that mother, you want to get the machine work with the rest of the data reinforcement. Learning is the machine eatery to continuously improve the outcome over time, the motions with zero in like a heat seeking missile and get closer and closer by the hour Reinforcement. Learning is very open ended. You're in person something. Was that one the machine to behave instead of living everything open to study and observation. You're giving the machine empirically call. Think of it this way, that when I was younger, there was a very simple video game called Bomb. The game had two separate bars on the opposite side of the screen. There was, like happened down as they tried to hit a ball to each side. In 2013 Google's Deepmind project experimented with call to see if they could teach a computer how to play. They set up a series of words with the computer everything. The computer hit the ball against Pattern before and every time that the opponent, Mr Poor, we've got another war I didn't pay against it still, and tried together as many rewards as possible. It only took a short while for the computer to start the master. The game a consistently beat human airs The Deepmind team is something called you learning this your learning help with some of the most complicated games and even more sophisticated wars, you're learning their city environmental stays. There are also possible actions that can respond to the states. Thank you. Learning you want the motion to improve. The quality of this is represented with the letter Q. You might have a kid next space invaders that requires you to shoot through Ellis to eventually win the game. This has a more complex see what system is not just a simple us happened. Your point is the ball. In this case you will start of the courtyard zero. Then you have the motion blur which actions improve the conditions. Each member in action put the state the cute girl from zero that you would go up based on the states and actions. This type of reinforcement learning is one of the most promising. Every machine learning. It allows the machine to go toe and dissimulation. Zoff actions and states until it tries the best strategy in 2015. Who goes? Did my project made the news when they're Obstacle program for Speed, an expert in the game Go half ago, You the phone unsupervised learning You don't manly by observing professionals playing the game. They're newer program op ago. Zero Real s family in Q learning. He didn't have to watch exports playing the game. Other Go zero simply went to the game and try different actions as a way to change is taking within hours this killer machine break off at the level that you must can't understand. In fact, after just 70 hours of changing offer, go zero b. The earlier divisional half ago in the personal 100 tribes, reinforcement, learning and specifically que learning allows the motion to quickly go beyond their understanding. It can help us get the steps required in a supervised learning. You didn't also observing and studying massive amounts of paper 10. Algorithms: Decision Trees: decision. Trees can be used for buying a reclassification. Charges of supervise mission burning Using dismantled, you can sit up predictors in nationalism. Outcome Decision. Trees are apartment Imagine. Apparently, you can show hot emotions making decisions back, creating a draft country for your presentation. Let's imagine that we're creating a decision tree to predict whether or not somebody is left to go to the page. Were created predictors. Let's wait a predictable told Sky, which because whether it's overcast when your son, then let's get another pretty quick all weekend. This predictable records, whether or not it's a weekend. Finally, let's get 1/3 critical that's called wind speed Hispanic with whether it's a windy day. You have two options. High or low. Now let's get one welcome variable. We'll give you the name. Let's quit. Yes, go to the page. It's abundant justification problem. So there are two possibilities you're classify whether yes goes to the beach where it doesn't go to the page, let's not creates a Gini data business on actually that were collected about the cash we could put status symbol Table of four columns. The first concert that predict person The Last column is the you could see the first UNOS of data in the training set. He pizza Sunday on the waking, and there's no wind speed. Daniels will go to the page if it's ready in a week there with having speed and you won't go to the page. This combination of predictors located didn't This is what we're doing good. Except it isn't really a decision to so great that you have to take ST are expected basin are predictors this just a pretty close Keiper? No, you should think of their own at the big chunk of the tree. We decided to use this site as the predictor, but you can begin with the predictors that you think makes thermal sands. Now let's without the next level about three. This is quoted this. You know you'll see old options in our room in these exactly options overcast. When you're sunny, it's at this table that you recorded percent of our house. Now remember that we want the Bonaly. We just want that yes or no cancer. Yes, go will not go to the big indigent decision or the total number of possible outcomes in the training data. There are four scenarios where yes, goes to the beach and see what he stays home this different out possible. Let us break the decision or down for several people. He knows this is the second set of predictors that are possible outcomes of the decision. No, you might know the family knows where the decision or this raining got me. That there are outcomes. Will it serve any day and you're still go to the page. You don't have to make a decision through any quarter, since there's no other option. Once a quick decision, Do you want to have a clear path when you said no? If you have trouble getting this, you might have too much and will be. This means that your place getting Mason is taking too long to get a yes or no answer. You can see some enter being the chief. When you look at the decisional problems here, you can see that there were three times when you went to the beach. When was high with into one was low. You're not really getting getting information here about whether or not just come to the beach To get me this anchovy. You might have to speed a data with a different predictor, or maybe not even use wind speed. That's a predictor to pass by whether you're going to be age. When we look at this decision to you get the critical idea about your behavior. That's one of the best things about decision trees. It's not difficult for people who don't know much about machine learning to get an idea of the different decisions. 11. Algorithms: K Means Clustering: another machine learning all within this. Kimmy's plastering this algorithm is after coupe is the skin nearest neighbours. But the only thing they have in common is that they both start with the later. Okay. Remember that genius neighbors is a supervised mission with you're classifying data based on what you're reading. On the other hand, Kim use cluster exam, supervise missionary with you, use it a great class there based on what emotion season the data. Think of it this way. Imagine what back of the any much of the the shutter has a lot Social Omar. All those get together in play. Those acting like people. They have a book prints and they chattered in God with each other. Each time they had a social hour, they were self organizing. Two different groups of friends now imagine that the shooter was closing. All the doors were going to be distributed into three different shelters across the city. The organisers of the animal shelter got together and decided to make it easier. The laws they were plastered them based on the stoops of progress. So the shelter dissented. 33 Castor's. That means that the cave in the came is equal three because we wanted to divide the groups into three plasters. Now imagine the machine learning all broken got started to start the machine Put her with yellow and blue color on random doors. Each color represented a potential class. They're based on the social group. This would build three century dogs. Now each of the century dogs will look at the mean distance between itself in all of this around in the house. Now the machine. You put this in color color on these dogs that were closest to the century parks. As you can imagine, this century dogs were selected randomly, So it's a pretty good chance that you won't have any good plasters. Maybe all three century dogs were in the same social cool. If that happens, and most of the ducks will have a very last distance between these three centuries, so the machine try over and over again until it beats the best injury. The home. It might even do this one pastor of the time at the end of the situation. The machine learning algorithm checks the various between each dog in this century. Once you have a good century, dog is pretty straightforward to put another does into each plaster. If you put on your talking toasters here, then you can tell social group. It ends up in just by measuring the distance from the century. Also, keep in mind that does themselves no class today to three girls. The money Piper. Six different social groups, but there are only three shelters of the machinery. Arquit has despairs to great plaster, the best represented off social looking. You should also want to make sure that you use Kimmy's plaster and give the doctor pretty supported the social robes. If the Doctor Chappy from book to book, there will be difficult for your pastors. This is sometimes called a hammerlock paper. Another challenges Kenyans is that we can be very sensitive talk liars. So if you talk, that's not really interested in hanging out with any of the other dogs who will still be clustered into one of these strict rules. So in a says that they will be forced to fight France, organizing dogs into treat plasters 40 different shelters is probably not the problem to running the day. But Kim's cluster is actually one of the most popular motion in our one of the more interesting application is when retailers discussed during police that gets promotions, then recreate re plaster that they call their customers somewhat have customers in those press shoppers. Then there create strategies to try to innovate somewhat past the most low customer. So they justify their little past amounts to participate in one of the programs. Many organizations are looking for better ways to plus up together their past. Amar's if we can get over our customers into one faster than you can really prove your business. 12. Algorithms: K Nearest Neighbor: in machine learning. One of the first Westerner about data is by classifying it with what we didn't know. You can group together a bunch of data business in your collecting these days because you already know this. Characteristics can classify the data using supervise. Mention learning. The Very Commons provides machine learning algorithm for multi class classification is Kenya Wrist Paper. This is an instant based, emotionally argument, and for what also called real easy learning. With lacy learning, the bulk of computation happens right before it. Classify your day. The learning doesn't happen continuously. Is that your on other computation? Wanting minsters? In a sense, you're sitting up on their energy for one big splash. Here's neighbors compare something you don't know so much already have. So in some ways they're getting immediately rewarded for the sizing party of your training data. The downside is that sticks out of computational power, So sometimes difficulty Is King Greece neighbor Some very last day, the stairs. Think of it this way. When I was younger, I used to work for an animal shelter. One of the most difficult job was trying to pacify the breed of each of our there are hundreds of different known that creates not only that, but the dogs aren't that close minded about who they breathe. So we have a lot of next three years. Each time we got in your dog, we have to put it up to several of existing dogs that were already classified. Then we'll look at some of the great three states. Maybe it was the shape of the face or the card of the hair. In a sense, the shutter was chanting, classified the unknown dog by looking for its nearest neighbor. Off course, it's not really easy to tell whether that was a Boston terrier were changeable. The closer the match, the more likely it was to because file, another way to look at it is you're trying to minimize the distance between the unknown dark and the no if the correct touristy so closely matched and was eventually distance between the undocumented neighbor, Maine is in the distance is a key part of Kerr's neighbor, the close in work? Um yes, neighbors, the more likely you are to be accurate. The most common went to do. This is to something called recruiting these things. This is a pretty sophisticated mathematical form. Rather can help see the distance between different data points now imagine had millions of dogs, and he wanted to classify them based on their real. The stock out may want to create a kicker various days, this one happy classified, though that share the same period. These are ethical predictors. So which is the way the neck of the hair? Now, let's say this to kicked a six and put them on an X Y axis by ground. Let's put the length of the hair around the Y axis and their weight on the X axes. Let's take 1000 classified dogs. No, Jane said. We'll put him on the ground based on weight and hair, no political, unknown dog and put in the same job. You can see that's not match with another dog, but it has a bunch of close neighbors. Let's see, we use a Q pie. That means we want to put a circle around our unclassified dog, and it's five closest neighbors. You can see that in the distance of the other dogs shorter. You probably get a much more active specification. Now let's look at the five closest game birth. You see that all of them are shepherds and told them our hostages, You can be somewhat confident to cast by your own on the business shaper. There's also a reasonable chance of the house key. Kenya's neighbor. It's a very common in powerful machinery. That's because you can do a lot more than just sort dogs. In fact, it's commonly used in finance to look for the best stocks and people predict future former's. 13. Algorithms: Naive Bayes: you've seen that suppose machine learning algorithms can pass for your data unsupervised mission. Running wild with instant plastic your data. Now let's look at an Internet, the principal algorithms based on conditional probability. See, you're looking to see how one thing impacts the probability of another thing happening. The most popular are good for this type of analysis. It's called the Beijing. Armed With based on the base unit, Statistics nine faces one of the most popular place emotionally confidence. It's going at you because it assumes that all predictors are independent from one another. Nine bases mostly used for Binali or multiply specification. So let's go back to our animals. Show there. Imagine that wanted to classify all the doors in the shoulder based on their different. Please remember that there are hundreds of different doctor years. On top of that most often makes your several Brill's This Look at this problem using 1/9 place motion to start this group three classes off the Braves. Will you stay of your Hans passport now, for each of these classes will have predictors, which is heading height and weight. Remember that some of the predictors will be closely auto correlated. But although it's more like behavior but nine place considered each of this predictors independently remember. That's why it's called you. Once your classes in predictor set up in the ninth place machine learning algorithms start with something called past Medical Probability. This is one of loss of each of these independent predictors and try to pay the probability that a dog belongs in each class. So let's see what happens when you drink with the department. The first, predictably look at his hair, the machine learning are good and convicted probability overdose with this hair land belong to any of district classes. It finds that although goddess Hedlund is a 40% chance of being a terrier, a 10% chance of being harmed and a 50% chance of the sport, the next thing you want to check is the one of those high again nine buyers. This inquiry Let the dogs here link with those Fine. It looks at this predictor independently and tries to calculate the probability that this unknown doubled teach class. So again it looks at the training data and find that there's a 20% chance that its interior , the 10% if it's a home and a 70% share villages poured down. The final thing you want to check is the unknown those way. This might seem like a strange predictor because it's closely related high. But remember, the nine pies is evaluating the probability of each particle independently. It looks at the training data in part that there's a 10% chance is a terrier of py percent chances a home in an 85% chance to sport down. So you now have this table with the unknown tops cast predicted probability. You can look at it. You can see that the dog is probably a sport. But remember, most machine learning problems are dealing with terabytes or even petabytes of the data, and it's trying to classify millions off. No, so to classify your dog, you don't want to use something called a wicked notification function. Since this is awaited function, the first thing you want to do is to be that way. It's, you know, decide which one of your predictors is the most addictive. You can pick this way by looking at the training data. You can also click it if it improves your accuracy his will is a weight off three for hair and for waiting time. So for the wicked multiplication function they predictor and multiplied. By your way, when you put up the predictors for the heading high one way you can see that the unknown dog is most likely sport. It's less likely that it's a terrier, very unhappy against harm. 14. Algorithms: Regression: gimmes cast a ring, and Kenya's neighbors are both instant baseball is learning. That means that you get other answers in one big splash. If something changes in your data, then you have to read on the algorithm. Scratch. This will make it difficult to scale the mothers because you're working with much more data on any one time. Sometimes you don't want instance. Based machine learning are quitters. Is that you want to say continues American relationship between different parts of a data. For that, you move you something like regression analysis. Regression analysis looks at the relationship between predictors in your article. Sometimes you hear predictors for people, parables, independent variables or even aggressors. Machine learning aggregation piled with them usually working this in the way. Once you have your training data, you make a prediction. They see how close you are to the outcome. Then you live it over in the work in and try to zoom in on the most accurate prediction of the outcome when I have a pretty good results. Take this training data and see positive test data. This is a supervised machine living with. You're taking the chilling data, leveling the quick output in your recently stable data with taste data. Vineyard vacation is one of the most common that's emotionally are with us with the navigation you want to create a straight time. That's what the relationship with your predictor in the outcome you want to see all the different data was close together around the straight line. Then you make the best prediction of the outcome. So let's think about how this might work. Imagine that the owner over that's been shot over the last year. You collected 20 days of sales data. Then you will know them porches away their data for those days. So with this training data, you create the simple X and y axes diagram Suszkin. This is when you better data on different parts of a diagram along the Y axis you so daily sales and along the X axis, he put the temperatures can, let's say, start the 60 degrees and 110 degrees. So you got a little different data about their background. Then you tried to quit the line of best here to do this acquittal and this piece of data. This is usually called the type of plane. Sometimes you also hear this. Call the trend line so you can see for most cattle. But I found that there's a very clear trend line. The more the temperature goes up the river, you're asking sales. You can also see that there's a few times that they have data pause. It will offset a train line. This could be because there was a past people or maybe just a lot of people wanted asking on a cool day. If you have a lot of data points that are far away from the trans time, that will be much more difficult for you to critical. Ask themselves has it stands. There are many out tires, so we can, usually the immigration to predict yourselves. So let's look at the point in this diagram. Let's say that all this week, the weather forecast says it will be running perfectly years. Now we can get their hyper being a trained them to drink from the correspondent gastric cells. You'll see that you can expect to sit around $3000 as you can imagine, not just asking. Shop owners are interested in using machine learning linear regression. In fact, if you have more data, it usually isn't making up with change. One interesting thing about the immigration is that there's some debate about what it's actually machine here. There's some truth to it because you think about it. The machine is not actually learning anything new about the day. That is, just using the data to create a standard statistical model. Is this about turning in one about protecting? Either way? Progression is a very popular way to try and predict future behavior, the keys to part of my predictors and look for some pop up in connection with Malcolm. 15. Problems That Utilize Machine Learning: in the HBO show Silicon Valley. One of the chapters use machine learning to go after the real problem. Here's is a smartphone to take a picture of some 40 pronto in the application tells him whether or not the heart he helpfully named application. He created application by building a small training, set off a lot of pictures and impressive millions of different put pictures online. Essentially, he was supervised, learning to go for more money specification even though there are many different supervised learning problem. The General foreign Tokyu categories. The first is a binary classification like not hot, the 20 plus classifications and their final immigration problems. He took this problem. General uses the same type of machine learning algorithms, not hard. Okay, says the mystical mentor to J. Crew ability that the image in the form contains a Honda. This is called banner classification. It's one of the most popular machine learning challenges. The banner reclassification. They're generally only two possible outcomes. Generally, yes or no. Is this hotel room come to be booked this week? Well, the stock market goes up, it's up north. Does this picture container harder? They have a pretty wife, won't cancer open Intensification gets a supervised learning. Remember that supervised learning the pattern label paper? Let me that someone teaches a computer a little bit about the right. In the wrong answer. You need to have a developer at the beginning to say the quite Devia you have to show the machine picked you with the Honda. Another common mission problem is multi pass specification. This is a problem where they're almost limitless. Possible categories. For this, you might imagine an application that we sent an image of a food fighting and classifies it based on several pretty pancreatic ways. Maybe we have something like a sandwich. Drink a big wars, then use a different set of statistical are written to try to put the data into these categories. You see some of these algorithms in later lectures. Another very interesting end of machine learning gets interrogation problems. Unlike finery and multiple specification, this problems tend to have a continuous solution him for chance and still trying to classify data. 80 didn't close. Think about whether it but there are more people. Not, for this is a partner reclassification because there are only two possible outcomes. That's useful, but what's really used for his know how murals elected to before. Here, there. Nope. Redefine our house. It's not the use of no question, he said. There are a range of possibilities that are more nous like. You can use statistics to predict that reporter 50 rooms would be a terrible like Bannon reclassification. Most recreation uses a formal supervised learning again. You're going to want to show the machine what it is for him to be entered for. Then you use something called linear regression. This diagram will show you trained that they can help you predict how many people will You're saying that the street problems will have very similar types of machine learning with us. The next step is to see how all of these different algorithms working Which one would be the best police challenge? 16. Choosing the Right Algorithm: something could have to do is the machinery specialist is to pick the best algorithm for your charge. In some cases, they're not going to have much of a choice. He thought they taste label. Then you probably don't want you supervised learning. Remember that label data helps understand both of you put in the outboard. So, you know, creating an application that helps suppressor home that you need a bunch of people data usual different times that happening by the data that's such a Z people square foot did your number of bath. Your machine doesn't have to pawn it. Some bathers, your device and able data you probably used unsupervised. Here. You let the machine create its own class those so you'll get the motion on the data that you have a different homes in the machine, decides with plus of make the most sands. Maybe the machine custom together off from the capital workability. It might even be a propia that's unknown. Once you have, the plasters will be able to extract some meaning. If you have a massive amount of on able data, then you probably won't use Kimmy's plastering or maybe some other way to have the machine class does. On the flip side, if you have a bunch of labeled data the removal used regression, unionist, neighbor or decision trees, you can also try a bunch of different our weapons and then take a closer look at their souls. Keep in mind that this can be talking to me again. You substantial computing power, so don't expect to get resulting media. Let's say that you're working with supervised machine learning. You want to use three different art. Britain's only training data decision Trees nine face Thank you, news neighbor. Then you can get the results and see which one hopes. High stable of practicing. You can also try something called Assemble Moderate. This is when he worked a grand people ensemble. So much learning part with us. There's a couple of points that you can create that samples this backing, boasting and stocking That was 20 quid. Several different versions of the motion money. Remember the decision trees keeping organized many different ways. You can't wait any different predictors for that would know. So if you want to use bagging, you just want to create several different trees and see which one of the best resolves you can also average other resulted circuit ecosystem outcomes. Boasting is one of several different emotionally armed with. To try and push the actress of your songs, you might rescue miscast oring in combinations. The Decision tree. You're taking the leaves off the tree and then letting them we should decide if there's any interesting booking. This is also a good example of semi supervised learning. Stacking is one of your several different machine learning algorithms and starting people. The accuracy they didn't want an in place prize is a form of stocking. It was called Future waiting in your stocking the greatest, several different predictive mothers in the stamping on top of each other. So if you could just talking this neighbor on top of 90 days, you told me I just had 0.1%. But over time they could be significantly improved. Someone is off the mission. Any competitions will start moving, 30 are. The key thing to remember is that you can think of each machine. Learning are with us a potential gold. You can experiment upon the best one, or you can work with several different posts as a whole people your accuracy 17. Fitting the Data: in machine learning will help to work with the training. Sad. This is a sports it off. The larger tape that used to 25 with in turn, These are Britains will help you create a mother that will work for the larger taste, Data said. Unfortunately, this training said mothers can not be a bit of a challenge. You can try to create a symbol this month will for small training set, but it's in principle when you're looking at larger data. This is the technical under painting. On the other hand, you could create a model that specs for enough to work with the data set but so complex is difficult to understand. This is typically call over painting. Let's think about what this might actually love by. Imagine work. Put a website that matches up buyers and sellers of policies you drink equated the mother that would have the fire spitting, developed the home. To do this, you create for main predictors. You use the square footage, the location, the number of bathrooms and the number of bedrooms you could use to provides machine learning and aggression. To try to pay some chance lies to predict the barrel one working. You notice that there's a lot of parents in the data. Sometimes this is quote boys. That means that there are a lot of passes with very different prices that have the same squint, portage, the same location and the same number of bathrooms. To fix this, you are more complexity. Maybe you create your predictors, such as quality, a few modern appliances and walk ability. Now, when you are more complex, you're also making the mother more sensible. That's because now they are minimal predictors to consider, but you're also making it more difficult to manage. It will be more difficult to see relationships between modern kitchens appears in the location. So here, over fitting the mother to the data on the flip side. Democratic, you think simple. You can create a nice trick patient chart that shows the relationship between the location of the house in the square footage. It will show that there is a close correlation between prices and big houses that are located in nice areas. This model benefits from being pretty simple. You would think that the big house in a nice area is more expensive than a small house in that area. It will also be easier to visualize. Unfortunately, this mother isn't very flexible. You could have a big, ugly, personal nice area. I also have a big house that hasn't been changing their kids. These problems was certainly reflect the pallor, but they wouldn't be captured in this model understands that houses will have a little theories, but my bias. They may well be clustered in the same place, but they will not be the right place. This is why it's called Under painting. You're under fitting good model to the data. It's not capturing enough information, so it makes an actual predictions, sometimes in data science and statistics time. Signaling Moore's signal is something you can use to make accurate predictions for the most , maybe just not for various in the data that cannot upper interview insides when you're working with machine learning, all good. But the trick is simply capture as much of the single as you can while not getting distracted. But there's no data 18. Following the Data: in the movie All the President's Men, the top informant of the Nixon scandal moving the dark parking garage, and told the report of all the money. Only by following the money could the reporter find the truth. I'm reminded of this trace when I think about machine learning all weapons, except this little following the money has to follow the data, but that's easier, said in. In fact, one of the biggest challenges in machine learning is figuring out you could get a mother is a bias sort of variance. Bias and variance are related, but there's still two separate challenges. Biases. The gap between your predicted value in the absolute power from period is when you're predicted. Paris are scattered all over the place. You can have a high bias in the low barriers. In that case, your predictions consistently wrong. But even worse, you can have a high pies in high previous. That's when you're consistently wrong, but in a very impulses that way, I think, will be this way a few months ago. I want my son to visit the summer camp over the course of a long weekend. They try to keep the parents a taste of the campers experience the very first morning that some because of the arteries to show us the ship they started give each of us a boy five panels. He devoted his toes. His five parents, which have it in the center of the target. His target physical example of low bison, all areas. All five of his arrows were getting tight in the center of the bull's eye. My wife was the second person to try. She never really shot a boy panel before, but she seemed to have a gift for it. She left most of Paris and had a satisfying park has depended on the target since she was a beginner. She didn't have the tactic. Group panels in the target is that you have a little bias. That means that in our school general together around the bull's eye. But she had high periods. It was one that stands straight in the bulls eye in the other four were getting in a circle . How is it there? Person to try? I went to summer, came myself, so I know people about how should the bone marrow. But I wasn't wearing my glasses, so I couldn't really see the target. I shot my heroes, but they had a high pies in the various different types together in the upper right hand corner of the target. I shot the areas consistently, but I was consistent with a bomb. My son was the final person to shoot at a target, and he had a difficult time pulling back on the ball. So he had a couple of years and important of the target. One of the key people get target in. Another one has been adjusting prom. He had a high bias in high various. His errors was better on the target in the equals, isn't it overshot undershot Tomorrow, a very similar thing happens when following the data, you can be like the instructor inconsistently. He don't target with their ability. You can also be like my wife wasn't a try but didn't have it figured out. With a little bit more environment, she could have consistently beat the market. It had me I was postseason, but I was consistent with a fall that means that the heavens still keeping on the variability low. But I was shooting in the wrong reaction. My son had a difficult time put it back on. The both of these areas are inconsistent hospital over the place. Remember that, vice in various ways of measuring the difference between your prediction in the they're not quite the wrong answers but different eyes that you need people with predictions. The key thing to keep in mind here is that you use different techniques based on whether you have too much pie, so much periods for me. All I needed to do was move my co people. It'll be done for my wife. She had the right target, you know, she was a factor. Predictions. And that's not the rule. When you know, sometimes you face you give it there. Expected to get happen. Parameters immigrated predictions. 19. Usual Challenges: as you can imagine what she's learning is still with thinking about challenges became watch for this tree. Common challenges when you're working on machine learning, being the first challenge is making sure that people in an organization can ask interesting questions. Next the Scepter in data separate from testing data, The final challenge is not to spend the most time choosing the right are with many emotionally specialists, working 90 organizations. This organizations will have groups of people really different scales. It might be difficult to make a connection between the needs of the business emotion, learning many departments. Why not have this chosen place last interesting questions? So one of the first challenges will be happy. The rest of the organization embrace someone respiratory mindset, money to post organization into asking interesting questions. I once worked for an organization that wanted to invest heavily in machine learning. They interviewed the local universities and hired the hospital's machine learning experts. This team immediately got were sitting at the technology they needed work with large data sets. Once that was also thought, they started to ask people what what the different questions that they wanted to answer. It would just feel shot off graduate school, so they were used by small date experiments on anything. Do it solves the business people and managers didn't happen. Experience asking with questions When we did, it was something like what people promotions customers like, which is something that would be so traditional database stools. The team feels like they built from my race car, and they were showing everyone how prepare that part. That's why you should get a good deal of time in the beginning, working with the business. What's more, interesting questions? Then, once again, this doesn't look good. First, yes, you can work with business to our creator. Another challenge more on in Poole is one thing. Doctor separation in data from the paper. Remember that training data is this amount of data that you set aside to build a modern, supervised learning. That's when I took this settings with commotion. Any so much learning things worked okay than extremely active mother religion. Then they should consult with their managers and the rest of the organization, and this is just make the mistakes off, mixing the training data back into the testing data. When this happens, you probably have much this active results. The first thing you want to keep in mind is that you should never makes a training data in testing data. This Mr Data, pretty supposed sections of our with in a sense, you only think you think your machine answers. Remember that working mothers Bijan, lies. That means that if you're mixing in some of your new day, that little old date that you're not giving the mother the first perspective, it needs to quickly taste its accuracy. You also want to make sure it's not overstating. Their captain is a model when you present great result of change. Date. I did my people physically or managers when the same mother this poorly later on. The best way to avoid this is to nothing. The presentation with training data wouldn't make a presentation. Make sure that you do it with unfamiliar data. That way, your manager several realistic view of its accuracy. The final thing to keep in mind is that it should be a concern if we have a strong bias toward any motion in our so much learning experts. You like some armed with better than fathers. If much, it's imperative we makes us there. Some algorithms are just like all songs that present families. Either way, don't be too concerned about making sure you're selecting The best are sometimes you want meto people toe with no store. 20. Next Steps: in this course you got an open field machine learning concepts and technologies you're seeing Hurricane you Several different emotionally are women's helpline for massively. That says. Then you saw how to mix and match this machine, learning with one Osama circling accuracy and it's all. Finally you saw some of the Children must learn again. How can be mysterious? Did you want counterparts? I hope that this course you get a better understanding of the different technologies used in motion burning his considerable potential behind this technology. Thanks for watching.