[MUSIC] Hi, this is Kence Anderson. We have an exciting topic for this video. Machine learning, algorithms can learn. And before we dive into this week's content, I want to commend you for making it this far. Of course completion statistics suggest that people who complete week three are likely, highly likely to complete the entire specialization. So good for you. Keep up the good work learning is a complex topic and the human mind is extraordinary far beyond any technology ever developed. But there are algorithms that generate a change in structure and behavior based on feedback. This is a simple form of learning in the field that innovates these algorithms is called machine learning. We've been talking about the skills gap, so let's imagine you're experienced employee has retired. The one with all those high value skills and you've hired a brand new operator that understands the process but has no experience in performing these high value skills. Getting this new operator up to speed is going to require significant practice. So how can you distill 30 years of experience into one month of practice? Well that's the exciting part. There's a technique called deep reinforcement learning. Where you can construct environments that will allow AI to practice the same way as the new operator would only much faster, much more practice in the same amount of time. And when you're autonomous AI system is trained, you can deploy it next to inexperienced operators giving them the benefit of skills acquired from years of practice. The way you bring it all together is with a technique called machine teaching, which is our focus this week. But first we need to understand the evolution of AI from machine learning and deep reinforcement learning to the game changer which is machine teaching. So first let's do a brief history of machine learning. I won't be talking about a full history of machine learning. I'll just be highlighting specific elements that are most important for you to have context on designing autonomous AI using machine teaching. So I want to take you back to 1966. So it's in the 60s and AI is very much in the popular culture. People are very excited about AI and its possibilities and the term robot has been coined and people are wondering what robots are going to do. And it's very much a hot moment in wave in AI. And at the Stanford research Institute in the bay area in California they built a robot called Shakey. So you might be thinking why do they call this robot Shakey. And we have a picture of him right here. Well Shakey didn't quite meet the popular or even scientific expectations of what a robot that explored its own environment could do. And actually he shook a lot and was pretty unstable which is where his nickname came from. But Shakey was the first mobile robot that could reason about its own actions as it explored the world. It actually used manuals, expert rules and expert system to make decisions. But, it lacked learned nuanced perception and that's where the need for machine learning comes in. That's what machine learning is essentially about. Machine learning is able to build complex correlations between variables. It can match patterns and it can build something like intuition about the difference between the inputs and the outputs. It's very much what robots and AI projects like Shakey lacked in the 60s. There's actually some really well known examples of machine learning models that we interact with almost every day. One of them is google translate. So when you go to google translate and type in a phrase and then pull down the language that you're starting from and the language that you want to be translated to and then you hit the translate button. It's actually a machine learning model. It's not a expert system, it's not looking up the definitions or looking up the translations, it's actually recalling or bringing up from. It's intuition based on the correlations that it's learned between for example, English and French, the French translation of what you just typed in English. Another example is computer vision. So when you have AI or models that can look at a video maybe on YouTube and then I can tell you whether what's in the video as a cat or a dog. That's computer vision based on machine learning perception. Another example is deepfakes. So a lot of us have seen is very popular and controversial actually, where you can take a person's face or image usually a famous person and then you can speak into a video. And the machine learning model will animate that person so the video looks like that person is saying what you were saying. So those deep fakes are based on machine learning models that are able to generate, it's called generative AI, those images. The other example I'll give you is Zillow and other real estate companies will provide predictions or estimates of what a particular home value, what a home is worth. And so when you go on to those websites and you get a prediction, it's actually a machine learning model using sometimes regressions. Using mathematical models and these machine learning correlations to tell you and predict what the home would be worth. All right, so what is machine learning good for? What should you use it for and when is it very useful? It's really good at fitting curves to data. So take a look at this data set on the screen. These data points represent the relationship between some input and output. We don't know what the input is. We don't know what the output is. It could be something from science or biology. It could be something from some real life application. We don't know what it is but we need to know a relationship between the input and the output. So each of these curves that are being drawn on the screen are different ways that you can use math to make a relationship or to show a relationship, make a correlation between the input and the output. The first is the most basic, it's called a linear aggression. It's where you take a straight line and you say well the relationship between the input and the output according to this data is loosely a straight line. There's others that could be exponential, it could be a logistic regression and what a regression is. It means that there's a curve, a specific mathematical line that describes the input and the output. Machine learning, which starts at these kind of regressions can build even more neurons, much more nuanced correlations between data than these models. The way that machine learning algorithms build these correlations is that they learn, they do something that's very similar to learning. They experience data, they experience more data, they experience more data and then they modify or change their perspective on the world in an analogous way to the way we're learning. I'm not saying that they learn exactly like humans and I'm not saying this is the human brain. What I am saying is that what machine learning algorithms do can imitate learning in a way that can provide a really interesting capabilities for autonomous AI. And because there's now algorithms that can learn, we're entering into a new era of machine intelligence. [MUSIC] The first era of machine intelligence was where the intelligence lived in the mechanism. In those days you had to build a new machine to provide intelligence for each task. For example the early mechanical calculators. There actually, there's a mechanism inside there's gears and other machine mechanisms that are actually making the computer make the decision. Whether it was adding or whether it's doing something like an automaton. There were these robot looking automatons that could write a word or that could wave hello or that could even ride a bike and a lot of times they would be showcased at the world's fairs. But what these automatons one, they were using a mechanical mechanism to provide the intelligence and to they had to have a different machine. You had to build a completely different mechanical mechanism in order to perform each task. Actually IBM International business machines, voting machine was like this. One of their first intelligence applications was a machine for counting votes. And it was a mechanical machine that provided intelligence to count votes for elections. Then about the time frame of World War two, Alan Turing came along and he said well, if you could have a machine that could accept different instructions, different sets of instructions. Like I could write a set of instructions, you could write a set of instructions and we could give it to that machine then you wouldn't have to build a different machine to do each different intelligent task. In this case the intelligence lives in the algorithm that set of instructions. So now you write a new algorithm to provide the intelligence for each task and that's what the touring computer did. And in fact that's the age that really in many ways we're still in because that's what software is a set of instructions that provides intelligence for a particular application. But now that algorithms can learn, I believe that we're moving into a different era, an era of teaching intelligence an era where the intelligence lives in the design of the system. So here's where you can teach each learner a new task. So in the same way that when my 8 year old learns a skill, maybe putting together something with Lego blocks or writing a five paragraph essay, I don't get a new child. I just teach my learner the new task and that I believe is the era that we're entering into with AI. We actually are starting to see this with famous AI like the Tesla autopilot AI which learns self driving, learns to pilot a self driving car or AlphaGo that learned to beat by practicing against itself. Some of the best players in the world at the ancient Chinese game of go. In some ways these are like savant learning algorithms but we're still having them focus on particular tasks. But machine teaching holds the potential for moving in the direction of more general intelligence. And in AI it's typically called artificial general intelligence and though that's not the goal of this course, it's important to notice that if algorithms can learn, it is possible to teach the same learning algorithm how to do different tasks. Using the machine Teaching framework in a similar way that providing algorithmic instructions prevents you from having to build a different machine every time you need a new intelligent task to be performed. So now that algorithms can learn, we're entering into this new era of teaching intelligence where intelligence lives in the design. [MUSIC]
