Machine Learning With Python Full Course 2026 | Python Machine Learning For Beginners | Simplilearn

Key Takeaways

Hey everyone, welcome to this course on machine learning using Python. Today machine learning is becoming a part of almost everything around us from movie recommendations, fraud detection to price prediction and smart assistant. Machine learning help systems learn from data and make better decisions. And that is exactly why this skill has become so important. Now it's no longer just for researchers or data scientist. It is now one of the most useful and practical skills for anyone who wants to work with data and intelligence systems. So in this course you will understand how machine learning works using Python which is one of the most popular language for building machine learning models. So we will start with the core concepts then we'll move into regression classification also understanding how to evaluate and improve models properly. So if you want to build a strong foundation in machine learning in a very simple and practical way, this course is a great place to begin with. So let's talk about what we have covered in today's video. First, we'll understand what machine learning is, how it learns from data, and how the basic workflow works using features, labels, training data, and test data. Next, we'll look at some important concepts like overfitting, underfitting, so you can understand why some models perform so well and others do not. Then we'll move into regression where you will be learning how models predict continuous values using techniques like linear regression, lasso ridge and polomial regression. And after that you will explore classification where you will understand how models predict categories using methods like logistic regression, KN&N decision trees and SVM. We'll also cover model evaluation using metrics like accuracy, confusion matrix and means squad error. And finally, you will see how to improve model performance using grid search, cross validation and pipelines. Also, if you are interested in mastering the future of technology, then the professional certificate course in generative AI and machine learning is the perfect opportunity for you. This is offered in collaboration with the ENIC Academy IT Kpur and it's an 11-month live interactive program providing you hands-on expertise in cutting edge areas like generative AI machine learning AI tools like chat GPT D2 and hugging face. You'll also be gaining practical experience to 15 plus projects, integrated labs and life master classes delivered by esteemed IT Kpur faculty. Alongside earning a prestigious certificate from IIT Kpur, you'll be receiving official Microsoft badges for Azure AI courses and career support through Simply Learn's job assist program. So what are you waiting for? Hurry up and enroll now. The course link is mentioned below. Now before getting started, here's a quick quiz question for you. Which type of machine learning problem is used to predict a continuous value like house price? Your options are classification, regression, clustering, or filtering? Let me know your answers in the comment section below. So without any further ado, let's get started. >> What machine learning is, which is a subset of artificial intelligence, right? That's uh basically um machines learning from data in order to uh make decisions essentially. Um so this was a big departure from the rules-based systems at the time, right, that were explicitly programmed to make decisions. So just think of an example like a really big kind of if this then that then that then that and and else if this this this right so bunch of rules that had to be pre-programmed in order to um come out with some final answer. Uh with machine learning it's the exact opposite of that. we're actually training something from examples from existing data um in order to predict something or um make some type of decision. Uh and so we're going to learn about the various ways we can do machine learning. But if you guys remember we um talked about some of this like the differences and the uh basically rules-based approaches to learning from data approach. Um and in included in that is going to be uh complex unstructured data. So things like images, text, audio. What handles those really well is uh deep learning which we will get to in the course after this. But uh those are certainly in there as learning from data even complex data. So we had this picture uh and I think this is kind of around where we left off last time was uh just distinguishing between those three terms. We see artificial intelligence, deep learning and machine learning kind of used interchangeably, but this is really how they fit in. Artificial intelligence is kind of a broad anything mimicking human intelligence. Um which doesn't have to be learning from data but uh machine learning is part of that. And then um one way to accomplish machine learning is to use neural nets which is the focus of uh deep learning. Um and so deep learning has been has found a lot of success especially recently with uh those complex data types like images, speech, text, right? So deep learning used all over the place. Even in um modern like generative AI, we see deep learning used quite a bit. Um it really anything that's using neural nets is uh going to be deep learning. Um and again we'll focus on that later but we're going to be mainly focused on machine learning for this course primarily machine learning that does not use neural networks. Okay so just models that are not necessarily neural networks be our focus. So in machine learning we had an example of a game uh essentially um learning what decisions to make uh based on the uh kind of current um state of the board. This could be a um you know machine learning example that uh learns from many previous examples. So a lot of data around these games are used to train these um kind of robots that can play these games and play them at a very high level. Um so there's been a lot of successes actually in machine learning and deep learning um around uh playing games like chess or go um using machine learning algorithms. So pretty cool. All right. So I think this is where we ended. We last time we said there's a bunch of different use cases for machine learning. So um recommendation system is going to be a big one and we will actually study that uh in one of our final lessons of this course. Um chat bots like generative AI doing sentiment analysis chat bots we'll study later but those are certainly an application of learning from data in order to uh generate responses to text prompts right. Um spam filtering that's a good example like classifying an email as spam or not spam. Um that that gets trained from examples and uh learning from data such as previous emails. Um social media posts analysis is another kind of text data um use case but you uh can do a lot with that text like you can predict the sentiment um you can predict uh the category of what what the post is talking about um those kind of things all can be done with machine learning >> and many other use cases not on this list that we will uh cover you know as we as as we go further. Okay, so this is where we kind of left off. Um, so what's doing all the hard work here is >> uh machine learning algorithms. So these are things that will um these are things that will learn from the data. So they are uh they they are basically um algorithms or sets of rules that uh or mathematical rules I should say not formal rules like in the in the sense of a rule system but mathematical um formulas and mathematical uh rules essentially that help us learn from the data. So they correlate the data to some type of outcome. So some type of prediction uh whether that's going to be as we will see whether that could be like a number like we're predicting a price or demand or sales um or it could be a category like is this transaction fraud or not fraud or what's the probability that this is fraud um so we have different kinds of predictions we can make with machine learning. Um but uh we will study the kind of the differences of those coming up. Um but machine learning algorithms are really what power they're kind of the models, right? They're the models that help power uh machine learning to actually learn from data. So we're going to spend a lot of time in this course studying those algorithms like the different models that we can build and what their differences are, what their strengths are, what their weaknesses are. we'll we'll learn a lot about those. Okay. So, I guess you can imagine like everything is so data dependent, right? Um we're learning from data. So, uh it makes sense that the quality of data really really matters here in determining how strong the model can be. Um so you see this graph here charting kind of the um high quality data um versus just uh any old data but a decent enough quantity of it. Um you can see that performance and the performance is measured by some evaluation metric. Um, so think of it as uh something like an accuracy. Like if we were predicting fraud or not fraud, how accurate can our model get at actually detecting fraud um it gets better and better and better the graph shows that the higher quality of data that we have. So there's kind of that there's a there's a saying in machine learning um called garbage in garbage out. What that means is if you have poor data, even the best model in the world, poor data is not going to result in having a good model that can be accurate and perform well. Um, so it needs to be high quality, meaning um there needs to be a decent amount of it and it needs to be labeled appropriately as we will will talk about. Um, and it needs to not have any, you know, significant outliers. it needs to be clean, not have those missing values, all of those things. Um, you can you have a good chance at deriving good predictions from higher quality data as this kind of shows. Okay. So, one thing we're going to learn um as we go along is quantity matters as well. So, not only quality, but a decent amount of it. And um we're going to learn those kind of rules of thumb like how much data do I need for certain algorithms. Um one thing that we will see is that uh the the basic machine learning models that we'll study don't need as much as a neural network would. It you know neural networks are going to require a lot more um than a basic machine learning model. So uh that's something we will see as we go along. But uh this is something we'll talk about and discuss with each model that we study is kind of how much data do we actually need to produce a high quality model. Okay, any questions uh so far? Okay, let's talk about the different types of machine learning that we're going to discuss. Prim, there's going to be two primary ones that we will study in this course and then a couple others that'll be a little bit more advanced that we won't get to but worth knowing about. Um, so there's going to be four total that we'll study or talk about and they'll be on this list here, which is um supervised learning and unsupervised learning. Now, I'd say the majority of our focus will probably be on supervised learning, and we'll talk about what that means, but we'll also cover unsupervised learning as well. And so, we'll look at the most popular techniques in each of these types of machine learning. Um, and then we'll talk about these two, but not really study them because they're more advanced topics. Um, that that will be beyond the scope of what we'll do. But uh these are going to be um different styles of machine learning that are going to be characterized by um what kinds of predictions they make, what kind of data they need and require. Um and uh what kind of outcomes they're actually producing. Um, so let's let's get into each of these, but uh the the one that we'll probably spend the majority of our time on is going to be supervised learning, but we will study unsupervised learning as well. We'll study both and we're going to talk about we're going to define both of those um coming up. And again, these will be a little bit more advanced topics that we won't spend too much time on. Um, but but we'll discuss their relevancy in machine learning. Um, and give a good definition to it. Okay. All right. Let's start with supervised learning. Now this is going to be uh a term that really refers to using examples. So using labeled examples. So here we say labeled data to help our model train. In other words, help our model be able to predict guided by specific input output pairs. So supervised really refers to the fact that we have answers. We have examples. We have answers with those and we use that collection of data to build our model off of so that we can predict um those kinds of things like a price, like a category, like a spam not spam. in this in this slide like we would be predicting if this shape is a square, a triangle or a circle. Um but but when we build a model for that, we have data that has an answer attached to it, right? We've talked about this before a little bit with labels. So there's a guide there that can guide us towards building our model. there's an actual every every example has an answer and that answer is really critical to help build our model off of. So, um that's it's almost like you have um a you have a bunch of exercises in let's say like a math textbook. You have a bunch of exercises and you have the answers and that way you can kind of check your work. you think about model training um that is the really a lot of that process of model training as we are going to discover is um basically checking our work against these answers in our data in our training data. Okay. So supervised learning is any type of machine learning that involves learning from labeled data in order to predict outcomes. Okay. Predict outcomes like now the the outcomes can be numerical. They can be like a price, temperature, demand, sales, revenue. They can be numerical, but they can also be categorical. So they can be like spam not spam, fraud, not fraud, cancer, not cancer. Um dog, cat, giraffe, those kind of categories. Um we could predict those. It's some type of outcome. Okay, some type of outcome. The key is we're using labeled examples to guide our model building. That's why it's called supervised learning. So we know in our data we know what the inputs are. Of course, those are going to be think of the inputs as like all of our columns and then we have a special label column that represents the output we're trying to predict. So if you think about that housing price data, the label could be the price and that's something we would build a model to predict, but we have answers for all of our examples in our rows. We have answers to help guide our model building. They help tweak our model because we know the answer ahead of time. So they're they're really good examples to build our model off of. Okay. So that's that's supervised learning. Uh in this example is circle not in the prediction because it's not part of the test data even though it's in the labeled data. Um no it just not necessarily. It just means that like we learn against all of these examples that have these answers and then when we observe new examples um we can try to predict what those would be based on what we've seen before. So I if there was a you know it's just a coincidence we only have two two examples in our test data like we could have a circle here in which case we would predict circle that's fine or at least we would hope our model would predict circle right that's what we're hoping may or may not get it right um but it's it's only not there because we only like we're just assuming that we only have two examples we're testing against but in reality we would probably do a lot more than two. It's just it's just a coincidence really. In reality, we would test against a lot more data. And we're actually going to see why we would do that. Like why would we train our model and then kind of use additional data um to to evaluate it? It's actually really important that we do that step to get a sense of how good our model is before we take it out in the real world. So if we apply our model that we build on our label data to um this kind of set of test data that we haven't been exposed to before. It helps give us a sense of how good is our model. So it's test is usually used for evaluation. So that's something that's something we'll study. How do we train? Uh it depends on the model. Um so training will be a sense uh will be an algorithm that will um basically update the model according to the data these labeled examples. Um every model is going to be different in exactly how it trains. So we're going to we're going to talk about that when we get to the individual models that we'll study. But uh loosely speaking, they're going to use the data to adjust itself. Like imagine adjust like tuning a bunch of knobs. Um, like the best example I can give you is we I think I did this one last week where you have kind of a function that predicts the price and let's say it has um weights like weight one with feature one, weight two with feature two, weight three with feature three. So imagine we had three input features and we we built an answer according to that. Essentially what we would do to train the model is adjust these um in order to get this correct based on our our labeled examples. Okay. So that's something we're going to learn about coming up shortly when we when we actually dive into model. Every model is going to be slightly different in how it trains, but at a high level it's going to use the training data with those examples, right? the labeled examples to help guide the formula essentially to adjust to generate the proper kind of model here. The these things are going to be adjusted according to the data in order to produce the correct output. So think about these as knobs that will turn. Okay. uh which type of machine learning is used? Uh probably supervised um which is what we're talking about now. So probably supervised because most people want to um build some type of model to predict something. Uh so yeah, I'd say I'd say supervise. Yes, we're we are definitely going to learn how to train. Yeah, we'll see. We'll do the code. Um, I'll tell you about how it's done. Yeah, we're definitely going to learn it. But what I was saying is it's kind of on a model by model basis. So, I want to wait till we get into the individual models, then we'll talk about how they're trained. But yeah, we'll we'll learn how to do that. But yeah, supervis is used all over the place. Even even for uh generative models, they use supervised learning because um like an LLM is going to use labeled examples in order to train, right? In order to train how to generate responses according to prompts. Um it needs to learn against a lot of text examples. So that supervised learning is what um results in that model, right? Learning from those labeled examples. Okay. It is yeah image image uh a lot of um yeah a lot of image processing is supervised like object detection. So the YOLO model is an object detection model. Yes. Um because it has to be trained right. It has to be trained on uh it has to be trained on images with labels such as this is what object is in this image. This is the box around the object. Um yes. So if if it's if it ever uses label data to train and build the model, it is supervised. So YOLO is definitely supervised and we actually we will we will cover the YOLO model later on in in our deep learning course. We talk about object detection. So we'll still we'll study that. But yeah, it's supervised Okay. So on the slide we have some common supervised learning algorithms that are we will study. So all of these we will study and understand what they do and how they work but just giving you some to name them. linear regression is kind of the one I just drew out which is the um this is the prototypical like easiest to understand model that is kind of the um exactly like this where we have a weight times a feature um a weight times a feature and then a weight times a feature and on and on and on. You can have as many as you want. um that is a linear regression. And so that is um that's a supervised model because we need this value here and we need all of our inputs in order to um actually train this model and generate all those weights um that that is uh that uses um labeled examples to help tune all those knobs. Um same with all these other models. So, we're going to talk about decision trees. We're going to talk about logistic regression and and SVMs, which are support vector machines. We'll talk about all of those, but they're all examples of supervised uh supervised learning. Okay, we'll talk about all of these. They're all supervised because they all require labeled examples in order to train them and and then subsequently use them. Okay. Okay. So what are some use case examples? So for for instance in uh supervised learning we may be predicting temperature based on yearly temperature trends. So we would have that yearly data as our um as our labeled examples and those would supervise the learning of a model that predicts temperature. Um, same thing with predicting crop yield based on um, seasonal crop quality changes. So maybe we have a bunch of features relating to crop quality. We could predict crop yield. Um, we would just need historical examples with those labels, right? What the crop yield is for each time period. Let's say we would just need those uh, supervised examples and we could easily build a model off of it. Um uh this this last one sorting waste based on known waste items and their corresponding waste types. Um that's kind of like spam. It's like filtering basically like a spam filtering. Um so think of it like the the shapes example. We sorting things into squares, circles, triangles. Um, same kind of idea here where we have a bunch of examples on what those um what those waste items should uh should belong to, like what waste bins they would go to, for example. Um, and those could be labeled and therefore then we could um understand what category of waste they belong to. Um, same thing with spam. something is fraud or not fraud, spam or not spam, cancer or not cancer. All of those are going to be supervised learning examples because they're going to require in order to train them, they're going to require data that has those labels. Okay? So, anything that has labels is going to be supervised learning. So again, this is where we will spend probably the the majority of our time is doing supervised learning problems, ones that we have labeled data. We're building a model and we're going to predict those those uh labels essentially. Okay, before we go to unsupervised, any questions about uh supervised. Okay. All right. So supervised requires labels in order to have an example to go off of to build your model. And that's because you're predicting those kind of outcomes like spam or not spam, cancer not cancer. Now unsupervised learning is completely different. It's the opposite. So unsupervised learning is where we do not use labels whatsoever. So we're not using any labels at all. So it's it it can be completely unlabeled or even if it's labeled, we're not using labels in any way. But um we primarily would say it's unlabelled data. We have no guidance because we're not using the labels in any way. We have no guidance to um predict anything. But that's because we're not really predicting anything in unsupervised learning. Generally, what we're doing is looking for some structure or pattern. Okay? Okay, with unsupervised learning, we're looking for some structure or pattern. So, um, one type of example that's very very popular is going to be this second one, which is, um, identification identification of user groups based on similarities or commonalities. Now, this is going to be a problem basically known as clustering, and it's a problem we will study quite a bit. there's going to turn out to be lots of different algorithms that can accomplish clustering. So what clustering attempts to do is basically say um we have data that's like this and then data over here and then data over here. Let's just group these together. So like this should be one group. This should be one group and this should be one group. And we can find those structures and say okay this is group one, this is group two and this is group three. one, two, three. And we can basically build what we would call clusters of data um based on how close together the points are kind of located in these kind of cluster zones like these boxes I've drawn. Okay. Now, that doesn't require any label to do which is really fascinating. So, unsupervised, you don't need any label at all to accomplish the algorithm. Um so, clustering is one good example. Um finding outliers or anomalies is another. So we don't necessarily have any label of what is an outlier or what is an anomaly. We are deriving that from the features alone. There's no guidance. There's no label um to doing like outlier detection or anomaly detection. Okay. So that's another good example. One that's not listed on here um but is also really important that we will study is something known as dimensionality reduction. So dim reduction and what that what this focuses on is basically compressing the data set a bit. So we take our data and basically compress it um so that but we do it in such a way that we retain as much information as we can. This is a very like smart compression and what it does is it lowers the dimension. Um dimension think of the dimension as like number of columns. Number of columns. So imagine we had 100 columns in a data frame. What we could do is actually reduce that down to 10. So like 10% of that. So we reduce it down to 10. And um but those 10 are it's not like we chopped out um 90 other columns. We um smartly kind of compressed all that information into these 10 new columns um that are compressed versions of the hundred that we used to have. Um so dimensionality reduction is is another unsupervised technique. requires no guidance, no label to do, but is um a really useful technique to reduce the size of your data if you're doing things with it. Um so this is another one that we will we'll study how to do it and basically more details behind it, what the algorithms are. Um we'll so probably those two in unsupervised will spend the most amount of time on clustering and dimensionality reduction. uh and supervised if some data is present but we didn't label it means in example we had circle triangle square in the training data we add pentagon but we didn't label that in that case uh yeah so every um in supervised learning every row, think about it as like every row in our data frame needs to have a label uh associated to it. It needs to have a a column that represents the label. So if we've never seen pentagon before, I can't use that as a label. So, it has to the Pentagon has to exist in the data if I'm going to be able to predict it, right? So, I can't predict, right? If we've never seen it before, we have no examples to go off. We have no guidance. So, how could we predict that? Right? We can't predict it. if it's if it's in there. So if if we have labels of Pentagon, let's say, then yeah, we could predict Pentagon. We could Remove. Remove what? We wouldn't if it was talking about the Pentagon. We wouldn't remove that. No, let me go back to that page. We wouldn't remove it. Um, it's just if it's not in our labels, we're not going to be able to predict it. So, Pentagon's a good example here. Uh, Pentagon is not one of our labels. So, it currently is not in our data set as one of the labels. We only have data that's either a triangle, circle, or a square. We don't have pentagon. So, I would never be able to predict pentagon. I'll never be able to do that if I haven't seen examples of it before. Okay. But let's say we had that in there. So, we had Pentagon. So, if we had Pentagon, um, we could have an example of it in our labels and then Yeah, we it could be then we could predict it. Yeah. Yeah. the the don't get worried don't worry about the test data. So the test data is just saying here's a new here's a shape what is it okay that's a square here's a shape what is it okay that's a triangle and we could have as many of those examples as we want in our test data so we could have a circle and say okay what's this should be circle right the test data can be whatever it whatever it wants but yeah if if we've never seen Pentagon before we're never going to be able to predict These are the the label data and labels are basically the talking about the same thing. The labels just mean what are the categories that are present in our data. So in this data we only have three labels that are present. So the labels is are relative to our label data, right? It's saying what labels, excuse me, what labels uh do we have in our data and we only have those three circle, triangle, square. So so Pentagon would not be part of those labels. We couldn't predict it. No. So unsupervised is not going to make a prediction. That's the big difference with unsupervised. They're not going to make a prediction like this. Um so unsupervised is not going to make a prediction. It's going to do something different like um basically say like these guys are similar, these are similar, these are similar, this is a cluster, this is a cluster, this is a cluster. It's not going to make a prediction. That's what supervised learning does. Clustering, yes, which is unsupervised. Yes, clustering does not require any labels. Unsupervised just means we don't have any labels. We don't require any labels. So the other thing unsupervised might do is it might say and again without the labels it might say that this is an outlier. It might say that this guy is an outlier because there's only there's only one of those and they're not like the other. So that that's something that um that's something that uh unsupervised could do. Um it it yeah and no. It kind of labels a cluster in the sense that um it would basically assign a number to it like this is cluster one. This is cluster two. This is cluster three. It'll assign a number to it, but it's not a very meaningful. It doesn't assign like a prediction label in in the traditional sense of a label. It does provide like a numerical index for the cluster to because what we want to know is like okay this guy has the cluster of one. This guy belongs to cluster one. This guy belongs to cluster one. This guy belongs to cluster two. This guy belongs to cluster two. Does that make sense? So there needs to be some like index of what cluster you belong to. So it's kind of like a label but not in the traditional like prediction sense. Okay. Very good. So again, unsupervised, no labels, you're doing things like identifying clusters, um identifying outliers, doing dimensionality reduction. These are all like structure and pattern oriented things. They're not predictions of a label. Okay, they're not which is what we would see in supervised learning. Okay, so an example would be that we take we put in the data um we can group together uh data such as images into categories based on similarities um which would be like those clusters. So there's no these would be groups that we don't have any label on ahead of time like we don't have we don't say that this image should belong to this this image should belong to this we derive that from the characteristics of the data. Um so think like a good example is um customer groups. So we would identify customers based on like okay do they have similar spending levels? How many days do they go shopping in a week? How much money do they spend? and we can kind of group together customers based on similar qualities. Clustering will find those groups that should exist. Um it will discover those groups based on um the similarities in the data, but there's no labels that that say like this person should be in this group, this person should be in this ahead of time. There's no labels of that. It gets derived during the algorithm. It's unsupervised. Right? There's no unsupervised really literally means no guidance. There's no guidance to doing it. We just derive that from the structure of the data which is the similarities. Okay. All right. So, a couple more for you. So, we had um supervised, which uses the labels. We have unsupervised, which uses no labels, looking for structure. And then we have something that's kind of in between which is um what is known as semiupervised learning. And this is where you use a combination of a little bit of labelled data, but most of your data is actually unlabeled data. Um and you try to get some use out of that label data in order to um build a model out of it. And so uh it uses the um it uses that label data to um generally provide some guidance on usually what happens with semi-supervised learning is you use your label data to kind of predict what the label should be for the unlabelled data and then you can go from there. So you can create artificial labels on this unlabeled data and then you can use all of it once it's all been labeled kind of like a supervised learning uh approach. So but but this is semi-supervised basically refers to the fact that you start out with most of your data not being labeled but you do have some labeled examples and what you can do is basically extrapolate those labels into the unlabeled data set and then provide some artificial labels and then now everything has a label you can do supervised learning. Okay. So, it falls kind of between um supervised and and unsupervised. Uh and there So, this is this is kind of rare. Most of the time you're not going to do that. You're actually just going to um prefer to just start with all label data. That's usually the preferred approach. Most of the time you'll actually just be doing supervised learning, not really semi-supervised learning. So, it's pretty rare, but um it it could like if Yeah, it could if the if we had a lot of examples of Pentagon and we wanted and so they were unlabeled and then we tried to guess what kind of shape they were um and provide an artificial label uh and then um then use that whole data set to build a model off of then then yeah, it could it could fall into this category. Okay, they Oh, going back to the question, they still use some kind of label data like age, gender. They use uh that's those aren't those aren't really labels. That's the features. So, yeah, they still use the core features of the data. They just don't have any like labels in the traditional sense of a label. Like you should think of a label as something we are trying to predict. So whether that's a price, whether that's like a category like spam, not spam, cancer, not cancer, it's something we'd be interested in kind of predicting. And so um in our data, we would have an answer for every row. We'd have one of our columns would be like the the result like the outcome answer that we're trying to predict. That's the label. So in unsupervised, we don't have any of the labels. We do have just the regular features like gender, age, income, square footage, bedrooms, bathrooms, all those things. Okay. So, we have semi-supervised that falls in between supervised. Now, the reason it falls between is be is because there's a decent amount of data that's unlabeled. In fact, a majority of it unlabeled. But what we can do is try to label it. We can try to take what we know from our existing labels and predict an artificial label and then use all that data together in kind of a supervised fashion for a model down the road. So that's kind of what this picture uh says is we can try to take um you know maybe we try to infer some labels based on we have some some labelled data here. We have most of our data is unlabeled and we try to supply some labels to it. Um like maybe we have a babies category of teens, a tween, uh you know youth and um adults. Um and then we try so we we take our our labels and we try to extrapolate those into artificial labels for this unlabelled data so that we can use it now because then everything has a label at this point and then we can just go ahead and do supervised learning from there. So we can do supervised from there. What we would prefer to do and what we'll do in this course um is just start with supervised. We'll just start with the labels. We won't try to derive artificial labels usually. We'll just start with labels. So one example in the real world is something like Google photos which um whenever you take a picture it can provide uh uh labels based on previous uh images in your library. So it can it can produce tags or um labels on those. Uh generally when you take that picture it's kind of unlabeled unless you go in and specifically provide some tags and some labels. But um if you don't do that it can still it can still uh make it can artificially create one of those based on the other label data that you already have. So that's um that's an example. Okay. All right. Last one in terms of machine learning. So we have supervised, we have unsupervised. Uh then we had semi-supervised which is somewhere in between a mixture of having some unlabelled data and label data. Um now we're going to talk about reinforcement learning which is completely different. Um it's it's completely different than the other three. It's a type of machine learning where we uh basically learn from interaction with the environment. And you might ask what are we learning? We are learning what actions to take in the environment. Um and the way we do that is by reinforcing positive actions that lead to a a reward. Um, so that's where the word reinforcement comes from is we we basically uh imagine like a child that's, you know, learning from trial and error. Like they're trying to crawl, they're trying to walk and they keep falling down. um eventually they learn how to do it through trial and error and they might get a reward or they might um reinforce some of those positive movements that lead them to walk or crawl um or they might learn from the penalties, right? They might learn from uh some type of feedback. So they might learn from falling down like, "Oh, that hurts. I should uh support myself a little bit better, right?" Or be a little more coordinated. Um and so they they learn from those actions and their interaction with the environment. Um uh so this is a complex um algorithm essentially uh it's it deals a lot with um again taking actions. Usually when you take an action something changes in the environment um then you kind of observe some type of feedback. So, think about like a a board game where you're trying to figure out what move you should make or another good example is like with a robot um trying to navigate a maze. So, like what route should it take? Should it move forward? Should it move backward? Should it move left or right? Those are different actions it can take. Also, like a self-driving car, should it should it turn? Should it speed up? Should it slow down? Those are all good examples of things that have been trained from reinforcement learning. Uh yeah. So, real world examples would be like in a board game, uh a a reward would be like if you win the game. Um or if you like capture a piece like in checkers or chess, that's a reward. A penalty would be like if you lose the game or lose one of your pieces, that could be a a penalty. um in a board game or sorry in like a a robot navigation task, it could get rewards for um moving in the right direction um towards the exit or like when it like let's say you wanted to train a robot on how to open the door and navigate a room. Um you would penalize it for bumping into the wall. Um you would give it a reward for moving usually oh like oh the algorithm themselves usually it's like a a step function um it's usually it's like a discrete function that kind of is based on the state so the reward it could be like um like depending on the let's let's go back to the board game example like the reward could be like or even the maze let's say like a navigating the maze like getting to this let's say this was the exit And this was the entrance. Then if they make it to here, they get a numerical like if they make it to the exit, they get a numerical reward of like plus 100, let's say. So it's just a number. And then if they uh like if they bump if they go into here, like let's say this is kind of like a death trap or like a pit, this this would be like a minus 100. So it could be like discrete numerical values could be the reward. If they're moving in the right direction like let's say we want to encourage going this way then we could give smaller intermediate rewards like this should be a plus like if you move forward this is a plus five this is a plus 10 this is a plus 15 if you're moving in the wrong direction away from the exit. Um that would be like a minus5 or a minus 10. Does that make sense? So they're they're numerical in nature and what you're trying to do is collect the most reward. You're trying to get the largest reward you can through trial and error. So you you try this out many many many times. You basically simulate running through this maze many many times. And what dictates it what dictates like where I should go is based on what I've observed in the past. It's almost like you're a child remembering like, okay, what move should I make from this space? Like, if I'm here, if I'm here, which way should I go? Should I go down? Should I go right? Should I go left? You kind of know that from experience. Does that make sense? Based on the reward that I've seen in the past, like when I've moved down, I've gotten a higher reward than moving left or right. That make sense? So, yeah, it's it's a numerical value as a reward. Yeah, that's a great question. Um, how does it differentiate rewards based on gain and loss i.e. chess? So it's it's a very comp complicated uh answer but essentially every so in the chess board you can think of the board as like every every um space is a state. So I could be in this state I could be in this state and then it's not not only is every every uh space but where all the other pieces are. So there's lots of states that are possible. Um, so the way there's a way to quantify essentially what's the value of taking a certain action like moving my piece left, moving it right, moving it up or down um given the rest of the state. So you're you're right, it may be beneficial to sacrifice. Um, but we would learn that through experience that okay, the best move in this situation is to sacrifice. We would we would have to learn that through trial and error many many many times which is to say like okay if I'm in this current state of the world right all these pieces are distributed in this way the best move for me right now in the long run to get the most reward in the long run is to actually sacrifice my piece and move it right move it into like a bad position theoretically but we know from experience that's actually the most long-term reward is from that position like moving it right may be the best action for me. So what you learn is how to take actions. And actions are usually like move right, move left, move up, move down. You think about like a self-driving car though, that's going to be like slow down, speed up, turn your wheel 10°. Um those kind of actions. So the the short answer is it's there's a calculation there that you learn what the long-term value of every state is every unique state and then you're trying to basically say what action should I take from that state given that current state of the world. Okay. And I really I really like reinforcement learning. It's actually probably my favorite field of machine learning. Unfortunately, we won't be covering it um in our main uh course. We have offered uh electives around reinforcement learning in the past. So, um stay tuned. Maybe when we get to the end of this program, uh we'll offer an elective on it and if enough people sign up for it, we'll we'll run it. But, um we it's not part of our we don't really cover reinforcement learning as part of our main topics. It's it is an advanced uh more advanced topic than than what we'll cover. But, um I I really enjoy it. Find it very fascinating. Okay. So, all of this is kind of um illustrating what I was saying, which is um you think of like uh the thing that's interacting in the environment like the robot or the car or the human moving a chest piece is known as the agent. It's interacting with the environment by taking actions which updates the state um of of the environment. So that's that's why you see this word state here. This gets updated constantly every time you take an action. Um ultimately what reinforcement learning is trying to do is learn the best action like what would be the best action to take. Um and the best action is is the one that leads to the most long-term reward. That's the best action. Um, so you have to uh you have to learn what you know what leads to a good reward by kind of experiencing this over and over and over through trial and error. So there's a lot of um kind of simulation or letting the robot try something a lot um in order to kind of learn what's rewarding and what's not. Think about it again like I think a good example is like with children, right? you kind of have to let them try things until they learn on their own what's what can they do and what can they not do what's the best actions right so reinforcement learning has made its way into other places so I I said like a good example is self-driving cars or ro robotics a lot of reinforce reinforcement learning is used there one place it's found its way into recently is recommendation systems have kind of merged with reinforcement learning Um, and this is because you you can imagine there's kind of a built-in reward for you clicking on a video and kind of watching it. Um, so that kind of reinforces that recommendation and then uh that's where um you can then kind of recommend a similar thing and see if that's rewarding and generates a click or generates some view time or watch time or whatever. Um so reinforcement learning has found its way into a lot of areas. Um recommendations being one of them because it's just natural for the idea of like what um should I recommend next to generate the most reward. In this case the reward is kind of correlated to did they click on it or not or did they how long did they watch for longer it's more rewarding. um those kind of things but uh place places where reinforcement learning have been used I said self-driving cars um games so uh one of the most famous examples if you want to look it up is the um Alph Go this was in 2016 um the Alph Go uh algorithm was a reinforcement learning bot that beat um some of the world's best Go players which go if you're unfamiliar or go is a um board game that is a little bit more uh complex than chess. It has more more uh it's a larger board um more pieces to it. Um but they there was a reinforcement learning powered bot that actually um learned how to play the game so effective it could beat um world kind of masters at the games was pretty amazing. Um that's the alpha go and that was by deep mind Google and deep mind in 2016. That was pretty that was only in 10 years ago not that long. Um so certain uh we said recommendation uh even autocorrect um learning to predict like what is the best correction uh to generate a reward which would be like you accept that correction or you reject it would be a penalty. Um so reinforced learning has been adapted to these kind of problems very successfully. Let's take a look at the packages that we will use throughout. So um of course we will rely on these three which we've already relied on to do a lot of things like numpy to do numerical manipulations and calculations. Uh mapplot lib to do any plotting and not only mapp but maybe seabour as well. both of those to do plotting. Um, pandas is a big one because that's where all of our data is going to be manipulated and prepped before it goes into modeling. So, all of that stuff we learned from pandis is definitely going to be applied here in this course uh as we actually build models. Um, so of course like these old ones that we've been working with quite a bit um still going to be useful here in the modeling stage. Um, mainly for different reasons though, mostly to get our data prepared to do some type of modeling or maybe to visualize it before we do modeling to get a sense of what it looks like, those kind of things. Um, sci is sometimes useful for certain uh um processing like in unsupervised learning. We'll actually use scyp a little bit to do dimensionality reduction or help us do that. Um so scypi will be used here and there and we've seen it before with hypothesis testing. We use scypi like the t test and z test came from there. Um some of the unsupervised learning stuff will come out of there. But the package we will use by far the most in this course is going to be scikitlearn which is here. Um and we've already seen a little bit about scikitlearn in terms of its pre-processing capability. So we use the uh minmax scaler and the standard scaler from there from the pre-processing module in scikitlearn. But it has um many different models built into it that we can use to help uh do our training and predictions. Um so it's a incredibly useful machine learning library. It is the industry standard machine learning library. Um if you're going to do anything in machine learning, it would be expected that you know how to use scikitlearn. Now what's really lucky about that is that scikitlearn is a really easy package to get used to. Nearly everything we do in scikitlearn will mostly follow the same pattern and so um the code will be extremely simple. They did a great jo