 So, it looks like my talk is going to be rather unique today in the fact that I would say it's more of a demonstration than a talk. And as a result, it does contain live coding, which is why I present to you, I wouldn't necessarily say a warning, but a very interesting comment on the fact that things could go wrong. It is live. It's a live demonstration. And it also involves me doing coding slightly under pressure. So bear with me. So yeah, point number one is going to be the fact that I am a technical specialist at Wolfram Research. Technical specialist is a rather interesting job title in the fact that it tells you absolutely nothing. Essentially, what my job is, is rather fun in that I get to play around with Wolfram technology, find out what it's capable of, and then present it to everyone else, which is why live coding. So without further ado, I'll go on to my talk as it's called AI Live Experience Not Required. I'm here to point out that it doesn't matter whether you're an expert or a person who's never even done programming before, I can have you doing programming in 20 minutes. And that's not a joke. I can have you doing producing yourself a machine learning classifier, deploy it, and use it in the form of an API or web form in the whole workflow should take you about five minutes. The training on either side, maybe a little bit longer. So without further ado, I'll get started. AI, one of the biggest buzzwords going around at the moment. What essentially actually is it? Well, my talk is both based mostly around what we refer to as machine learning, which is a subset of AI, AI being the automated use of algorithms chosen for you by a computer. All of you here will have done mathematics at school. Absolutely all of you. One of my favorite and possibly most embarrassing points on my mathematics at school was trying to do a line of best fit when I was younger. You'll have been given a pencil and a ruler and told to draw a line of best fit through the points, trying to make it as even on both sides as possible. Like so. If there are any mathematicians in the room, please don't shoot me. I know I'm doing this rather haphazardly. There's a reason for it. If you're like myself, you probably would have been told off for trying to use a pen instead for when I made mistakes. However, unlike me, you will have gone through your mathematical education. I did go through mathematical education, but didn't get told off for that bit. Where you will have learned various automatic methods for determining a line of best fit for your data. Up to the point where if you're lucky enough to get through a mathematical degree, you'll in fact get to this point where you'll be able to, you'll be given and taught a method whereby you can find the perfect line of best fit to describe your data. Anyone out there who actually does data science on a day-to-day basis will be able to tell you that's useless in a practical sense. It's great. It tells you exactly what the data is you've got. However, if you are trying to use this for prediction or analysis, as soon as you go outside those existing data points, your results are useless. So why does this relate to machine learning on AI? Well, that is just the automatic choice of method to provide you a fit to your data, whether it be a predictive model or a model for classification. That's all machine learning and AI actually is in the background. So what does it actually give you? Well, this is the bit where I'm hoping to make all you guys laugh. So here we go. As I said, live demonstration. So to start with, I'm going to go for image contents. I'm going to take a live feed from my webcam, put it through a machine learning classifier, and it's going to tell you what it thinks it sees. Thankfully, I'm a person. And yes, I'm lifting my arms up to prove that it is live. I can dance if you want. Preferably, I'm not going to. And you can see that it believes I'm a person. In fact, I have this interesting clicker here. We'll find out what this believes it is. But not clear. I can see that one. Breathalyzer, flash bulb, autoloader, cell phone. I can agree with cell phone. It's a bit better. But this is a live demonstration showing you that image recognition is really simple, to the point where we can change it to image background. It believes I'm in a bamboo forest. Bit worried about that jail cell one, but we'll come back to that. Movie theater indoor. I like that one better. It's a bit better. There we go. How about facial gender? I really like this one. Because it's even going to be very nice to me, in which case I will highlight I am male. However, occasionally, if I put my hair forward, it suddenly thinks I'm female. Which is the reason why I'm a technical presenter with long hair. There is a use for it. And it's, in fact, frozen slightly. There we go. So machine learning is brilliant, but it's not perfect. And it's not just limited to images. Whilst images are the most amusing, I'm not going to do age or gender or facial expression, because they're just downright insulting. But we can do things such as imputing missing data. So taking a neural network, providing it with an image, and getting suddenly a colored image of a dog. Whereby you can see it's colored in the grass for me, and including the two little flowers hidden in the background that you probably couldn't even see originally. I can quite happily tell you, if you try this with the London Tower Bridge, it actually works really well. In fact, it looks slightly better than the real thing, but I'm not meant to say that. Or, if that's not impressive enough, how about taking that original black and white image, not only giving it color, but providing it with depth. So we're turning a two-dimensional image into kind of a three-dimensional image. Come on, PC. Please don't let me down now. It's thinking. There we go. So we now have a 3D depth-based image of that dog, AI Live. I was kidding. So I keep boasting that I can teach any of you guys to do this kind of thing without much training. And I'm not kidding. So much so that I'm willing to put my money where my mouth is, so to speak. And now it's not happy with me. Come on. No. Told you it's live. There we go. It went and crashed on me. That's because I was playing around with it earlier, but hey, that's my fault. What I'm about to show you is the fact that, whilst I can show you all these cool little funky demonstrations, and they all seem to work nicely, he says, I can also show you with one or two simple lines of code, so much so that you can actually read them. I can teach you to do things such as supervised learning, which is the kind of thing you do with a child. You show them an apple. They know it's an apple. Or unsupervised learning, where we just throw a load of data at the computer and see what it comes out with. Warning, dog classifier, because I saw the cat one earlier. So to start with, classification. The identifying of an object or item you have in front of you should be nice and simple. Humans do it all the time. Therefore, training a computer to do it should be just as easy. So I have two examples. The first one is just going to be classifying numbers into categories. Anything higher than number three is class B. Anything lower is class A. Rather boring, but provides a point. You'll see that my coding, if you can call it coding, is pretty simple. These are lines of code. The first one is me setting up my day training data set. The second one is creating a classifier function and then testing that classifier function on the number 1.2. As I said, if I give it three, it suddenly changes to B. It's easy to do. That doesn't really provide you guys with anything to take away, apart from, hey, look, the code's simple. Which is why I have an image example, because I'm a very visual person myself and prefer to keep things visual as if possible. Here is a load of images. I've taken off my favorite search engine of day and night. You can search this yourself, nice and easy to do. I drag and drop them into the notebook. No playing around fancy dancing things to put the information into the notebook, just drag and drop it. Put the arrow in and go day and night. Create a classifier function. There we go. It's doing the automatic AI interesting side of things where it's choosing a method for me. I'm not having to sit here and specify all the kind of things you normally would have to do as an AI expert. And now I can straight away use it. That took all of three seconds, maybe 15. I've done that live in front of you guys with no magic or no hidden requirements. And already I'm able to classify six day and night images. Can anyone spot the issue? It's deliberate. It has misclassified the second image, the one with the large 0% finance thing out of it on the car. It's a dark image. How would you be able to tell if it's day or night? If all I've given it is dark images for night and daytime images all have bright blue skies in the sun. If I taught a child that way, they'd make the same mistake. So AI, whilst amazing and absolutely brilliant, will make the same mistakes that humans can do. That's a slide I've left in for interest purposes, because I will be providing copies of this talk afterwards so you can actually go in and play around with the examples, et cetera, yourself. And this is just a bit of information on the different methods that that classify function could access. These few here, these six, nearest neighbors, logistic regression, support vector machine, you can read the rest, are just some of the ones that the predict function, the classify function I just used in front of you guys may choose to use. In fact, if we go back to it, we can see, by clicking the little plus, that in this case, it chose logistic regression. That was automated, as I said. I didn't choose that. The system did it for me. It's how you can do AI and machine learning in about five minutes. Doesn't matter how complex it is. This is where I come on to the point of machine learning. Does everybody in this room understand how it works? If I gave you a particular complex neural network that does object recognition and classification, would you be able to pick that apart, explain to me what every single different layer does, what the different weightings mean, and provide me an exact example of what it would give when I provide you an unknown image? No, let's be honest. No one in this room should be able to do that because this kind of stuff is incredibly complex. It's amazing. For that reason. So you don't necessarily ask how or why did it work. You just need to know, as users, does it. In which case, I can provide you numerous examples of yes, AI and machine learning work. This one is a classification of irises, a particular type of flower. They come in three particular varieties, satoza, versicolor, and virginica. This is a very simple numerical, example. One thing to remember with machine learning, if you provide it an example it's seen before, it will cheat, it knows the answer. So what I have to do is split the training data up, 100 examples, retrieve the other sets of data as my training, as my test data, check they don't intersect, they're completely separate, and then set up my classification. This is where, once my classifier is created, I can test all 148 additional examples in just a few seconds to provide you guys with statistics on just how accurate, in this case, sorry, it's only 50 examples, on just how accurate that neural network on that classifier function has been. There's the accuracy, 0.96 or 96%. I can give you a confusion matrix plot so you can understand which ones it got wrong. In this case I'm aware that's a bit small, or if you wish I can show you what the best classified examples are, or even all the properties you can query for that classifier function. So I can imagine some of you guys in this audience are much better experts than me, that kind of information being able to provide it to you automatically, really quickly, will mean that you can prove to me, or I can prove to you, that yes, this machine learning works. So I've shown you some fun examples so far, proved to you that you can do it in a few seconds, and shown you that if you're an expert, or you just want to know if everything works, you can find out, again, in moments. You can do the same thing with prediction. We're not just limited to classifying anymore. If you want to set up a series of data, and create a predictor on that, you can do. Again, it takes a few seconds, in fact, that was four seconds, apparently. I don't believe it. Got a bit of GPU acceleration going on in here, that's why the computer looks so space age. But then I can use that predictor function straight away. I can guarantee you now I could sit there with that predictor function for ages, typing in numbers, and I'll never understand what it's doing. However, I can query it for a distribution to find out what that prediction looks like. In this case, it's a normal distribution with those parameters. Or, better yet, I can plot it and show you exactly what that predictor will give you. Because I'm a visual person, I love seeing actual outputs. Or, I could show you this on a real-world example. At the moment, a bit of personal history, I am trying to buy a house. It's interesting and difficult. Worth it. Here I have a series of information on Boston, because our main office is over in America, simple as. We have an example data, which I'm just gonna quickly pull, just to show you that it's a series of numbers that each represent these kind of statistics. Apparently, these are factors in Boston that seriously affect house price. Well, you'll notice that some of them are a bit unusual. There's a reason for that, and I'll show it to tell you in a second. But again, creating a predictor, simple as wrapping, predict around it. Nothing complex as usual. That one's gonna take five seconds this time and it's going for linear regression. Who knew? Now it's changed to gradient-bruster trees. Still not sure what they do. Anyway, again, predictor measurements in a few seconds. Have it go through and test 168 different examples. You didn't even see that update. It was that fast. I hope. But already, I could give you a comparison plot, which, as I said, provides us with a visual example of what actually happened in the background. There is a very faint line here, which shows exactly what the prediction is expecting in terms of house pricing. Anybody here who has anything to do with houses automatically knows that house pricing is not linear. In fact, anyone who wants to highlight that these ones here are inaccurate is correct. I mentioned this to somebody earlier. It's something called artificial stupidity. This system is amazing. It's predicted accurately, goodness knows nearly 168 different points to within an acceptable standard deviation, which in this case is 3.94. The bits that have been inaccurate are the result of missing data. This system is amazingly accurate, but it can't predict things like people's interest in houses because a celebrity has lived there, or does it have a swimming pool, which makes it more desirable, apparently? Trust me, if you live in England, having a swimming pool is not such a big thing. You could just hop out into your local pond, they're big enough, and it rains all the time, so they're always full. But it's this kind of human nature that is somewhat predictable, but if you don't provide the computer, the AI, that information, it's not gonna be able to factor it in, which is where AI and machine learning tend to make mistakes. It's pretty obvious when you think about it. I can quite happily tell you that the worst predicted example, oh, that's different for once, is this point here. That's not, is that one? Somewhere around here. Nope, 23, 24. Surprise. It's a live-in example, apparently I'm wrong. Normally it's this one up here, because there was a factor to that house which made it particularly desirable for humans, which was not one of these. It's the reason why they're slightly abstract in bits of information, because they're not all options for changing house prices. So I can classify, I can do prediction, I can apparently get internal self-testing errors and crash my computer. But these are, again, this is not what we're limited to. You can do unsupervised learning as well, which is where some amazing stuff comes in. If you ever hear a machine learning in AI being referred to as magic, as experts, you all believe that to be false. And it is true, it's not actually magic, it's just unbelievable. To the such a case that the average user has no idea what the machine learning is actually capable of. Leave it to load for a second. The reason I mention this is because I can provide this PC, live in front of you guys, a series of three sets of dog images. And without telling the system anything about those images, I can actually get it to group them up into the different breeds automatically. And should this continue, I will be able to show you. Have I got something running in the background, maybe? Well, I think that's conclusively a no. I'll try an older version. I did warn you it was live coding, so anything can happen. In fact, one time I spent five minutes being insulted by my PC because it was telling me I was 43. Not that that's relevant now. It's the reason why I skipped over the age of demonstration at the beginning. So yeah. Whilst these things are impressive, they can be very offensive if given incorrect information. I apologize. This one's slightly faster. He says, is it now freezes? You've got to love live demonstrations. Sometimes they can go perfectly, sometimes they can go wrong. And considering I'm doing machine learning on this PC as opposed to doing anything cloud based or anything like that, it means that if any of the hardware is running slightly hot, suddenly I have problems. Here we go. We're running hot again. Come on. All right. So before I go on to the dogs, I'm going to quickly show you some sequence prediction. Predicting what a person can do can sometimes be particularly easy. Quick demonstration. Step forwards, step forwards. What am I going to do? Step forwards. Three steps forward. It's pretty easy to predict that because you can see what I'm going to do. So again, doing that kind of thing with a sequence of one, two, three, four, you can pretty easily predict what the next number's going to be without any particular effort whatsoever. If I give you the numbers one and two, what's the next number? Three. I did hear someone say it, but you're not being particularly interactive with me today. I'm not sure if you're all asleep or you're all waiting for lunch. We'll find out, won't we? Give you the numbers one and two, it gives you three. If I give it one and three, it automatically goes to four because that is the next number, even if you skipped a step. Sequence prediction tends to be used for things like finance, for trying to predict future trades and things like that. But I want to give you a bit of a slightly different example. What if you were to give it the entire text of a book like Alice in Wonderland and create a sequence prediction from that? How do you predict sequences from text? Well, what you can actually do is if you've ever tried using the predictive text on your phone, there's a fun thing I tried doing previously, where if you type in a letter and then tap the most suggested thing on your phone that it predicts, sometimes it will give you a few sentences or just a few words that make sense and then it will go into gibberish. Well, I can quite happily tell you now that if you do this on your PC with a sequence predictor, you actually get something a lot more accurate. So here is the system's prediction of what would come next after I give it the words Alice was thinking. Alice was thinking of. That's English, it makes sense. Alice was thinking I. Well, maybe she thinks I'm an idiot. I am stood up in front of you guys trying to entertain you at the same time as show you coding. Well, Alice was thinking about. They're all sensible predictions and it's all been done on text that I've just given to the system. Alice was thinking while, Alice was thinking and. Alice was thinking that. They're all sensible suggestions. If I get Alice was thinking, let's try about. What's it gonna go for next? About in, about it, about a, about her. Sensible suggestions apart from the in and the it. Maybe the in. But you can see that sequence prediction on text can be really simple. This case, it's gone from nominal sequence, whatever that is. Which brings me on to two advanced issues which I just have to warn you about. Just in case you do intend to use machine learning and you've not necessarily done it before. My first one is unbalanced data. It is the most common problem in machine learning. Is the fact that if you don't provide a system with enough data for one example or another, it's not necessarily gonna predict that outcome or classify that outcome as accurate as it, accurate as it probably should. This here, for example, is a series of data from our office on a particular day. You can see it's been purposely chosen so there is only one female. So what happens if I predict, if I ask that classify function, the exact height of that female, it should classify them as female because that's the only example it has. I told you, if you show something, if you show the system something it's seen before, it knows the correct answer, right? In this case, not so much. No matter what I try, that is the probability of ever being classified female in our office on that particular day. It never goes above 0.40%, which means that the classifier will never predict you as female. Apologies, ladies. Why is this? There isn't enough data. Statistically, you are more likely to be male than female so it's gone for male at all times. Using something called class priors, though, you can adjust for that. And you can say, I'm sorry, that's not quite true. There is a 50-50 chance of being male or female. I would hope that's the case in here, in which case it's now adjusted its percentages appropriately. It now says there's a 64% chance that you are male if your height is 1.6 meters on that particular day. I think you all agree, that's a bit better. It's one of the ways of handling unbalanced data, and it's one of where my first warning comes from. Never falsify your data, or aim your data collection specifically to achieve desired results. You should be doing your balancing and your adjustments for bias afterwards, like I'm doing here with the class priors. Because all you're doing is providing that classifier a particular understanding, which is not true. It's better to adjust them appropriately afterwards than if your adjustment, it turns out, to be whilst idealistic but false, the system will still adjust for it appropriately. And one other unbalanced issue, one other advanced issue is unbalanced outcomes. There is a common sci-fi understanding that once computers get intelligence enough, they will rule the world and we will all be killed. Thankfully, that's, goodness knows how far away, simply because computers, whilst they are intelligent, they kind of need to be told what to learn and what to do next. They're not quite there yet. And one of the reasons for this is the unbalanced outcomes. Let's say you're trying to treat someone with a severe illness, that if left untreated, we'll kill them. Normally the big C is the example that gets used here. Let's say that if you provide an individual treatment and they're unwell, then great, they should get better and they should survive. However, if you provide a healthy individual that same treatment, they're now just as likely to die as the person who doesn't receive treatment who is ill. That's the unbalanced outcome. Under no circumstances do you want to provide that medication or treatment to that healthy person because you're going to kill them. Machine learning, whilst amazing, doesn't understand this kind of system. It really does not understand it. All it will do is classify a person as healthy or diseased and determine whether or not they should retrieve treatment. If it understands that receiving treatment makes you more likely to survive, you're always going to get the treatment because it thinks you're more likely to survive. You've not told it that it's going to kill them, but you can use something called a utility function to adjust for that. It's a bit like a reward system. If you get it correct, here's a point. If you get it incorrect, minus one. If the person is diseased and you classify them as healthy, then that's the minus one because that's really bad. You're going to kill them. If the person is healthy and you've classified them as diseased, it's not as bad because you haven't treated them and you're not going to kill them, but hey, we'll see. Basically, you can adjust for this kind of thing, and this is where the human interaction is required. I don't know if you've heard on the news, you've seen on various news websites about the fact that they're trying to use machine learning and AI to assist in cases in court. Machine learning and AI can quite happily classify whether or not someone is more likely to go on to offend or not, but it won't tell you that as a result of them going in jail, they've reevaluated their life choices and have realized this is a really, really bad idea. My family's now going to suffer because of me. How are you going to tell a machine and AI classify that? That's going to take you a long time to figure out, which is why we need human interaction at the end of this. It should be using an advisory tool, not an automatic controlling system. It's the reason why everything we do at Wolfram, we never link it straight up to the end result, straight up to some kind of action mechanism. You need some human to read the output, understand what's going on, and then act upon it. Now I can go on to my dog's example, a slightly more fun one rather than possibly killing people. It all comes down, this magical machine learning and AI, to something called feature extraction. I will admit, I stood up on stage, you all did feature extraction within a few seconds. You will have taken a look at me. I've got hair, long hair, glasses, dark t-shirt, probably look slightly nerdy, probably an IT guy. Not going to blame you if that was the case. I sat down at the center parks in the UK and the guy straight away said, you work in IT, I went, yes, I do, and I'm proud of it. That's feature extraction. It's something you guys do every second without thinking about it. You've read that slide, you've picked out the important information you've ignored the rest, and then looked at me waving around on the stage and probably focused hopefully back on me. The clever thing about machine learning is this feature extraction. The ability to extract information from an image or data automatically without any human deciding that is the important thing you need to look at. That is why machine learning suddenly took its exciting new steps, even though it's been around for ages. When someone realized you could get a system to understand which features are important, that's when machine learning suddenly improved because before we were telling it, oh no, look, if you're doing facial recognition, you need to find the nose, you need to find the eyes, the ears, and then you need to match it to another picture where you do the same thing and you compare the two. Facial recognition at that point had a horrible accuracy and no one could understand why. We turned off the bit where we were suddenly telling it which features to look for and already the accuracy doubled. To start with, you need to understand what the system's doing in the background. In order for a machine learning algorithm to use your data in whatever format it comes in, it needs to convert it into numbers. Simple. Here I have the work in a large cat and I'm converting it into a series of numbers just so you can see it. Capital A is a number 36. A space is a number three. You can understand that. Simple so far. If I convert that into tokens or words, again, it comes up with three numbers which you guys can roughly recognize and possibly use later. This feature extraction, by picking out the important bits, means you can ignore all the random rubbish in the background, also referred to normally as noise. So if I create a classifier that recognizes the words cat and dog in a sentence, I can ignore things which are like uppercase characters and diacritics, which means that if I happen to mistype this is a doog, then it still recognizes the fact that I'm referring to a dog, despite the fact that I mistyped you. That's why you gotta be careful with having the multiple languages on one keyboard. Don't ask. Interesting project. If you then provide this to the machine learning example, so I'm just pulling in from an open source database. I can't even remember how many examples. I think it's 100 of each dog breed, Chihuahua, Bassett and Labrador Retriever. Basically I'm holding as many as with a memory on my PC can handle. And I could show you a typical set of eight examples. Just to point out, these are not clean images in any way. I've done this on purpose. Some of these include other dogs of different breeds. They're trainers, including people hugging dogs or one of them's got a dog with a huge massive branch which can't get through a door, which is actually really quite funny. Anyway, the idea is the images aren't clean because in reality when you start trying to classify dogs, the image is never gonna be clean really. Let's be honest with you. This one's got a teddy in it. This one's not even looking in the right direction. And I'm no idea what he's doing. But when I turn that same feature extraction capability onto those dog images, what you'll find out is that it gives me a series of numbers which mean absolutely nothing to me. This is what I was referring to when we turned off the which features you actually need to look for. In fact, that's taking a long time to download the neural network required. Well the rest of my example runs on that so I'm gonna have to wait now. But the idea is that it does translate that simple picture into a series of 49 different numbers. I have no idea what those numbers represent. They could be how circular the nose is, they could be how wet the tongue is, how curvy the ears are. But those are just things that I think it might be looking for from my understanding. The feature extraction's so complex, as I said, we've turned off telling it which one to look for that it is amazing by doing this all on its own. It can ignore the human cuddling the dog or the random Chihuahua sat next to the Labrador or whatever. This is why it's absolutely incredible and also why my example hinges on this. Yep, I'll have to wait for that one. The idea I'm trying to show you is, as I said earlier, I'm going to feed in all the images of the dogs and using something called feature distance which is just how close are these two images in the feature space of those 49 different measurements. You can actually get it to automatically cluster the groups of dogs together by breed without you having to tell it anything. There we go. There's the vector that represents the dog, this particular dog. If anybody at any point can take this example, take it home and figure out what all those numbers represent, please let me know because you're amazing and probably one in a million. So that would be incredible. Or I can just show you feature distance whereby I can provide it with two different images of dogs and it can tell you how far apart they are. So these two dogs are more closely related than these two. I know by looking at the images, you can do that yourself in seconds. But remember, we're trying to go on about AI here. So without further ado, I'll provide it all those images of the dogs, hands off, I'm not doing anything. You can just see feature space plot of those images. And it will provide you a diagram showing you the clustering of those dog breeds without me telling it anything. It doesn't understand anything apart from the fact that they're images, okay? I haven't told it they're dogs. I haven't told it the image size or anything else. Here we have the dark colored Labradors together. Light colored Labradors, including one with a, ah, that's clever. Sorry, I've not seen that before. There's one here with a light colored Labrador and a Bassett Hound. And funny enough, that's probably the reason why all the Bassett Hounds are down here. Because it's seen one in the image and it's classified them all together. They're all Bassett Hounds, absolutely all of them, including that hidden one there. All your light colored Labradors are up here. Next to your two hours, which are also light colored, including the one with a scarf. That's kind of cute. I've provided the system a load of images and it's clustered them for me. Try doing that with images on, with data from social media. And suddenly it will show you how certain groups of people cluster together naturally and how they share their information. Or maybe try doing it with customer data and it will show you how certain products are related to certain customers just by putting in the name of the data or the product alongside the customer data. These are things I've actually done as part of my job and you can get some surprising information out of just automatic clustering like this. I've done very little work already. But it took me a few seconds at most after the internet download did the thing. If you particularly want to, you can get into neural networks. We have a framework for neural networks whereby we've set up all the different layers. You can combine them and construct them however you want. You can access out of the box existing neural networks from our repository. Where we've got neural networks for object recognition, classification, you name it. They're all ready to go. Feature extraction, image processing, regression, speech recognition, video coming soon, you name it. I don't really care. I'm just here to show you that it's easy to use. Examples of how to use this stuff and access it straight away. But if you want to sit there and combine, retrain existing neural networks or adapt them to your own uses. One, two, three, four lines of code and you can do it. I'm really not kidding when I try to tell you guys this stuff is easy. Anybody can do it. If you want to, and you're an expert and you really like doing this kind of thing, you can retrieve a neural network and provide it an image just as a use. It can identify a tiger. Or it could take the first five layers and show you what it's doing inside. At layer five, that's what the neural network is doing. Any ideas? Anyone understand that? Okay, either that's a blanket no or no one wants to interact with me. I'm gonna go with a blanket no. That is actually an image representing each of the vectors that the neural network has come back with. The idea being that white is a positive match and black is a negative match, as in I'm not interested in that. What they're looking for, no idea. Again, if anyone can understand that, please come back to me and let me know. I really want to talk to you because I've spent ages staring at this kind of screen and understanding none of it. So I've gone through some supervised learning. I've hopefully entertained you guys and I've gone through some unsupervised learning where I've shown you plenty of pictures of dogs. Feel free to change that to anything you want. Otherwise, that's the end of my talk. If you do have any questions, let me know. And if you want a copy of the talk, just email me. You can have a copy of it, you can do the examples yourself, turn it to whatever you want. Thank you.