Introduction to Machine Learning Methods (Prof. Steve L. Brunton)

again I just want to point out this is something I'll talk about more in next lecture most of the tasks we have in general in engineering can be written as optimization tasks okay model reduction you're trying to find the best model in some metric of so you know at some reduced rank our to describe your data control is usually posed as an optimization problem sensing sensor placement actuator placement closure models these are all optimization problems the thing that brings them together at least in fluids and in a lot of the systems of interest to modern machine learning is that their high dimensional nonlinear non convex and multi scale so these are optimization problems that twenty thirty years ago we couldn't solve even ten years ago and with emerging techniques with better data bigger computers and better optimization algorithms we can start solving these problems okay good so I'm gonna just walk through types of machine learning examples of what's easy and hard in machine learning then we'll talk about neural networks how you can model dynamical systems if I have time we might talk about some historical context of kind of understanding what's happening today and machine learning with respect to what has happened in the history of science okay so I'm gonna start here with types of machine learning so we have some complex system this could be you know maybe I have a self-driving car and there's a camera viewing the world in front of this car maybe this is my fluid dynamic experiment or simulation it's just some system that's generating data hopefully tremendous amounts of data but not always okay and so the idea is that you're going to have some data that you're going to use to train a model and there's a lot of different ways of slicing these machine learning algorithms I'm picking one categorization which is whether or not you have labeled truth data okay so oftentimes I when I collect my data I know what type of data is so I'm taking pictures and if facebook has all of your pictures it got you to label them right oftentimes in the past like you would click on who's in the image and you would type who that person is and you would label that data so Facebook would know who was in that picture okay so that's labeled data maybe you run a Windtunnel experiment for a new airfoil you measure the lift coefficient okay that's labeled data unlabeled data it would be more like the self-driving car example where you just have a camera and things are happening and you don't have someone expertly labeling what's happening okay and so if you have labeled data if your data has tags telling you the features that you care about then that's supervised learning makes sense and there can be continuous or discrete labels so for example if I'm trying to build a classifier to tell pictures of dogs from pictures of cats there are only two labels dogs and cats category one and category two that would be a discrete classification and there's lots and lots of methods for a discrete classification support vector machines decision trees neural networks so on and so forth if you have continuous data like the lift coefficient example then this would be a regression problem okay so all kinds of different regression models the the Sindhi model I told you about yesterday is a regression model Gaussian process models are regression models and so lots and lots to choose from and one thing I'm going to point out a few times I'll probably say this more than once is that I want us to not get confused machine learning is not just neural networks and don't let people tell you otherwise there are going to be people when you say you do you know machine learning or you have data they want you to train a neural network and that is you know the only worldview they have on machine learning that's not what machine learning is berrin yeah I'm sorry so so what I want to do is I want to separate tasks from how we solved these tasks none of these tasks are going to be build a neural network a neural network is a tool that you can use to classify it is a tool you can use to build regression models there are many others ok so another categorization if you have unlabeled data so you do not know what data you're seeing this is kind of the older one of the older branches of machine learning is unsupervised learning okay so this is classically called data mining you have a bunch of data you're trying to find patterns and extract meaning from this data but no one tells you what you're looking for so this is like searching for a needle in a haystack again you can have discrete or continuous discrete unsupervised learning is called clustering you're trying to find groups of data that behaves similarly to each other and differently between groups okay and you can also have continuous unsupervised learning which generally I'm gonna call embeddings or dimensionality reduction so p OD proper orthogonal decomposition and principal components often fit into this kind of unsupervised world okay and I'm only listing a very few of the of the methods here for example ten years ago people didn't use neural networks for classification much they weren't the best Amazon used decision trees random forests support vector machines those are the industry standard methods yes if your labels are discrete it's a classification task I mean whether or not your labels are discrete or continuous so my question was if I measure lift for example in a discrete time so I don't mean discrete time or continuous time I mean whether or not your labels are discrete if I'm only trying to classify between pictures of cats and pictures of dogs there are two distinct labels if I'm trying to classify or predict what the lift will be on an airfoil at an instant that can be any range of a continuum of values for the lift coefficient that is not a classification that's a regression task let me point out also thing I say so uh maybe I should just say in general a third of what you hear in machine learning is going to be at the very minimum and so you get the fun job this morning of trying to figure out which third of this if I get 80% of the way there I'm pretty happy I'm gonna oversimplify that's my my job today is to oversimplify things for you okay good so the middle block I think is particularly interesting this is where a lot of the most cutting-edge research is happening in machine learning because and there's actually a really interesting dichotomy between the successes of machine learning and industry think Facebook Google Amazon they have mountains of labeled training data and they have billions of dollars literally billions of dollars worth of labeled data that they got us to label for free okay it's much harder as scientists for us to collect that label with data that costs money okay and so a lot of work is going into what's called semi supervised learning where you have limited labels you have an expert kind of semi supervising your process and it might be too expensive to label all of the data so you label 1% of your data ok and there's a few different kind of big categories here I'm classifying them based on whether or not you want to model your data or actively modify your data whether or not you are a free agent that can change the world or whether you're purely a passive observer in the world and so some of the really exciting ones are these generative models the Mona Lisa picture I was showing earlier that's a generative model where you basically have to kind of in this case neural networks or learning algorithms that are fighting with each other one is trying to trick the other one and so it's semi supervisors in the sense that the neural network itself is supervising another neural network ok reinforcement learning is a really interesting class of optimization where you are actively trying to change how you interact with an environment so you can think about the game of chess ok you are learning the rules you're learning your opponent and you're trying to design a control strategy to in some sense Maksim some objective function okay and so you can also then take what you're kind of your model of the system and make decisions that then change the system okay that's a pretty interesting perspective and that's something we should all be thinking about because we're not just going to be classifying images we're trying to manipulate our system and you can also do this with with fully supervised learning with just generally optimizing and controlling your flows okay good okay so that was like about as shallow as you can get in terms of classifying machine learning algorithms again I want to point out there are tasks there are the quality and quantity of your data and then there are methods of actually solving the optimization probably when you're trying to write down and those are three different things okay so I when I when I talked to like executives about machine learning I always give them this little spiel on what's easy medium and hard because that's probably the hardest thing for people to figure out when they're just learning about machine learning you see all of these examples and it's not obvious what is possible and what's not possible what would take one afternoon and Tenzer flow and what would take five years with a concerted research team and millions of dollars okay so this nice xkcd maybe some of you have seen this so telling whether the photo contains a bird this was actually pretty old because it's now easy this was hard a few years ago labeling what's inside of an image but you know if your data has a geotag it's pretty damn easy to tell where it came from okay like it's all about what's in your data and how hard your optimization problem is okay so image recognition facial recognition this is one of the original successes of machine learning this is kind of why honestly that's why we're in this room today it's why in the last ten years so much attention has been given to this is because finally machines can classify images about as well as humans can maybe not in every circumstance maybe not as reliable robustly but it really is approaching human level performance Facebook can tell if two people are the same or different about as well as you can okay pretty remarkable interestingly they don't do it by just taking two-dimensional images that gets you about two below 95% classification accuracy but humans are up in upwards of 97 98 percent ability to tell who's the same and who's different any idea is how Facebook cleared that last 3% which is always the hardest sorry okay so that's a great one is they have some idea some prior of who you would possibly be taking pictures of based on your network that's that's definitely true but they do this test on so the same different test is for complete strangers people you don't know can you tell if they're the same person or a different person so what would get Facebook up to that 97 98 percent accuracy and again this is five years ago so yes so they're already using color depth so they do a few things right off the bat the first thing they do is they will take your face and there's pretty good algorithms to find corners of the eye corners of the mouth and corners of the nose and what they'll do is they'll stretch your face onto a regular grid because everyone has different geometry and you can't compare people's faces if my eye is where your nose is so you have to stretch everyone into a common grid that's step one they've been doing that for decade step two they actually estimate the geometry that the structure of your skeleton the three-dimensional spatial geometry that's how they cleared that last 3% it's a little creepy they're modeling your skull so you see this picture here you as a human have a pretty good idea that that is a three-dimensional nose that casts a shadow and that's extra information that you can use to classify okay so that's this additional understanding of the world that that was very very difficult to learn but that's what got them to the to clear the gap okay so image classification is a huge deal this is just a tiny tiny tiny excerpt from the image net database so how many of you have heard of an image net image net was this I forget how many 10mm labeled images hand labelled images something like 10,000 categories of images with thousands of examples for each like a thousand pictures of broccoli labeled broccoli a thousand pictures of tractors labeled tractor humans did this this is one of the most impressive datasets collected and that large data set was what allowed these new neural networks to be trained with such high high accuracy so again machine learning does not start with a neural network it starts with a mountain of quality data usually ok you can zoom in and see there's all kinds of stuff in this and there's thousands of examples of everything ok and so image classification 15 years ago is very difficult now it's almost trivial we take it for granted ok you can hold your phone up practically and classify an image this one is really cool this one actually fooled me I thought I gave this talk a few times before I realized that these were actually misclassified like these are generated images that were generated to trick a classifier so these are pictures of my broccoli my tractor some purple flower these are generated images from that same statistical distribution I don't know if you can even tell the difference between some of these tractors if you're sitting in the back maybe you can tell that this broccoli looks really not edible but it's kind of amazing that you can now generate images that's pretty cool ok but it's not perfect this is the cutest blueberry muffin I've ever seen I mean it's not difficult for you to tell the difference right everyone in this room I see some people squinting everyone in this room can tell the difference between a blueberry muffin and a chihuahua okay my sister just got a fried-chicken dog he's adorable but this actually still is challenging because machine learning algorithms build models based on the features that they have they don't have legs and arms and a body to go carry them through the world and build experience they don't have they do have ancestors now but they don't have you know generations of knowledge and modified neural networks like we do we have an understanding of the world that goes much deeper than what our simple machine learning classification algorithms understand the world in now these algorithms are being deployed at scale this is another thing you have to think about is just because you can demonstrate that something is possible can you make it robust can you make it streaming real-time scalable so now you know with self-driving cars and surveillance these image classifications are running concurrently in hundreds of thousands of platforms all across the world right now ok so it's a little scary but that's the world we live in is that this is now ubiquitous this was pretty cool image captioning was for a long time thought to be impossible for for machine learning algorithms and I think within the time I've been giving this talk this became possible so this is image captioning you can give Google a picture of a scene not just a piece of broccoli or a tractor but a guy playing a guitar and it will caption that image for you pretty neat of course it's never perfect this is probably my favorite one this looks like where I'm from in West Texas and you can see that there is no horse but it's it's pretty remarkable and the the performance here is surprisingly surprisingly high okay Atari games this one is really neat you can use these reinforcement learning algorithms to build models of whatever virtual world they're given data so this is a google deepmind reinforcement learning algorithm they're given the pixels in classic Atari games okay so just like when we were kids we played games until you know we couldn't see straight anymore so it is Google now and this is block out and you can see this is literally it just learning how to play the game it has some idea that what it wants to do is make these numbers big okay so someone tells it what the objective is and now it knows that it has some control input and it's just gonna futz around until it can figure out what to what to do okay after 2 hours of training it's actually quite impressive it's playing better than I could it's a very simple game this is its entire world its entire existence lives inside of this box so that's a little sad but something amazing is gonna happen so after two hours that learn to play but after four hours something really really special happens it learns the trick that like only what happens yeah you're gonna have to wait this is the suspense this is so I've been making fun of machines for years and I think they know good let me skip to the actual part of the movie so you don't have to watch this again okay good okay so it's playing like an expert and after four hours you're going to see the transition so remember when you were a kid and I don't know this was before the internet when I was a kid or before everyone had it there was one kid in the neighborhood who could just destroy everyone okay that's this google deepmind look what it's doing it's gonna find a weakness it's gonna dig a hole come on there you go and now it's salt like the game plays itself did you see that it's amazing it learned something I mean this is a simple world where you can there's one trick and it learned it in four hours it's pretty impressive now there's actually a really great paper I forget if it was a science paper or not but you can download this original Atari deep learning min paper and there's a list of which Atari games it could totally destroy and which ones totally destroyed it and it's really interesting to start thinking does everyone remember asteroids simplest game ever you turn right you turn left you you know fire fire you know I try to destroy asteroids it could not play that to save its life for some reason I suspect it's because of rotational and translational invariance which is really hard for for machine learning okay but it's not all about playing games you can also teach robot machines how to do human tasks so this is the catch in a cup task right you swing a ball on a string and you try to catch it in a cup so when we were kids we this game and this robot you can see it's trying it has to have a white background and a high-resolution camera trained on it at all times to learn anything after 15 trials after 25 trials close probably can't untangle itself but this is about how long it would take a kid like if you had a 10 year old trying this it wouldn't take them you know it would take them between you know fifty and a hundred trials probably there you go they learned how to catch now this is again this is its entire world it's not doing anything else except for this but you know and and this is interesting it requires some knowledge of physics I believe they probably baked in some rudimentary laws so that it could learn the physics it needed to this one I think is pretty fun this is a video my friend and Lisowski showed me it's a DARPA video just showing how advanced our robots have become I mean robotics is really hard we take for granted how seamlessly we walk this wasn't that long ago this is 2015 so ed likes to joke that in the robot uprising all you have to do is hide in the bathroom but you don't have to lock the door yep so it's still quite challenging yeah so this is actually a really interesting point so so my friend Nathan could choose a version of this which is the Boston Dynamics robot everyone's seen the Boston Dynamics big dog incredibly impressive four-legged robot they also have their Terminator looking robot the two-legged super big terrifying robot um we're talking about how these robots are now you know it's inevitable that they'll be used for be weaponized and so in general everyone knows about the Boston Dynamics robots Google bot Boston Dynamics sold it right back why I would argue and this is a bit you know speculative it's because they're generic machine learning didn't work at all for Boston Dynamics robots those are it's a it's a case study they're self-driving cars a lot of this is built on on machine learning and deep neural networks the Boston Dynamics robot is a physics-based hand-tuned model with control you cannot just give it a bunch of you know data and learn how to move a robot it did not work ok very interesting so we do have a lot of robotic capabilities but it is not using the same technology like Google sold it back and you can you can make your own conclusions why they didn't find it profitable to keep the Boston Dynamics and it asks it forces us to ask the question what is the next thing that will be possible okay so wondering what was difficult ten years ago is oftentimes not difficult now and so it makes us wonder what are the thing that are truly unique to a human okay and this is actually quite relevant for the young people in the audience if there is something that everybody can do I think at some point we will train robots to do something similar not always that's that's that's a broad generalization like for example walking and catching and picking up an egg is easy for humans and it's still very hard for robots but eventually I think they'll figure this out but things that require real expertise where it is not something that everyone can do not everyone can write the next great American novel I don't think machine learning will ever write the next great American novel I would love to be wrong but I don't think that will ever happen okay and so you have to think about what is unique and requires an understanding that you can only get through a unique lifetime okay machine learning is never going to be able to build those capabilities it might be able to make a American Novel there's actually a really fun machine learning paper where they generated Harry Potter fanfiction it's not too bad it gets pretty bad starts off good I mean you know in Hagrid's hut there were disdainful shrieks of his own furniture you could totally imagine that happening in Harry Potter okay it's really interesting some of the texture they get really well but their sentences don't all make sense okay there was a group of MIT that basically created a paper generator okay and they got a bunch of papers accepted and made you know fools out of a lot of leaning peer-reviewed journals I was just seeing on Twitter a couple of days ago someone generated again for all successful ICML papers they basically generated a network trained on all of the thousand successful papers that got accepted and they can now generate their own papers you know pseudo figures and pseudo titles and things like that so we you know like you can create things that are remarkably good at fooling someone who's not paying any attention at all but there are some things that are fundamentally difficult right so we live in the world of uncertainty and chaos and dynamics right so yesterday we talked about how non-linearity can give rise to this sensitive dependence on initial conditions it is very very difficult to predict where a hurricane is gonna go with models with machine learning from the data we have there's only so much information and the uncertainty you have amplifies in time so there are fundamental limits to what information you can extract from your data okay I like this one I think this is really kind of a fun example so for a while Amazon would suggest what else you might want to put in your basket okay and so if you bought whatever this person bought it might suggest that you want ball bearings magnesium and gunpowder or you know in this case I think this is this is the material for making thermite aluminum and magnesium okay so it thinks that whoever is purchasing whatever they're purchasing also wants to make thermite and it would suggest you know shrapnel bombs and thermite and all kinds of things is very embarrassing because and it's just scary because this means someone was actually like this is not Amazon just making this up someone was actually putting these together in carts enough that this was in the database okay so you know it can go wrong and I think this is interesting for us to think about what could go wrong what could go wrong with you know a pervasive machine intelligence everywhere guiding and suggesting things for us so for example I saw a really interesting talk from a professor at Harvard who took legal case studies from from the city of New York over decades and decided that they would train machine learning algorithms to basically Robo sentence people okay so the judge gets a docket and they get some suggestion based on okay well I've seen this case before this age group this crime this record you know blah blah blah what is the chance that they'll repeat offender chance that they'll become violent and so on and so forth and they get to decide on what sentence is a define is that you know do they set the bail do they have to go to jail things like that and so you can build a machine learning algorithm to basically automate that and give a judge a packet that just shows them the distribution shows you where this person is and gives a suggestion about what they should do we have to decide whether or not we are comfortable with that and what are the moral implications of that what they found what was interesting about the talk was this was basically an example of why you should never do this because if you have a historically prejudiced system you are going to bake in prejudice into all future decisions okay we have to be worried and careful and cautious about this and this is a new slide I just added this morning because this wasn't around the last time I gave this talk this is you know pretty brand new with these generative models you can take a static image in this case of the Mona Lisa and you can transfer that over to I mean this one is just incredibly expressive right you can you can now generate fake movies from from like one image that's that's impressive so it's constantly a moving target you know every day there's a hundred new papers in machine learning every year there is at least one major advance like this something we didn't think was possible that is now possible and so you know you have to be thinking about what is what's easy what's medium what's hard in machine learning when you're applying this okay but there's a big difference of course between industry and their resources and their motivations and our resources and motivations as scientists so for example if Google wants to train a network for some task that it you know spends billions of dollars on if it gets 0.1% improvement in its performance that translates to a tremendous amount of money saved and so when they I mean a lot of these these companies when they show you their brand new natural language algorithm or their brand new image classification algorithm what they don't tell you is that they had a wall of graphics processor units maybe 10,000 maybe a hundred thousand and they train the same model with slightly different parameters on every single one of them and they cherry picked the one that gave them 0.1 percent better performance cross validated because they don't give a about interpretability they just want it to work a little bit better and make that more money and that's not usually our goal as scientists and we don't have those resources we don't have that capability it's a different completely different optimization where we live and I don't think that they often tell you kind of where they're sitting in their optimization landscape they'll make their code open-source you have to train it on a hundred thousand GPUs for two months to to reproduce their their results okay okay good any questions yes so I think there is also something which is very important is to point out there are some function evaluations lose up to output is not important so if you like if they if ms amazon tells you that you have to buy termite you don't care I mean you know I mean if even if even if it's nothing I mean even if it's a wrong suggestion the impact of this wrong evaluation is not important but if you design an airplane and folds this function evaluation is important this is such an important point I hope everyone heard this there are there's a there's a whole scale there's there's many shades of gray about what decisions matter and how much they matter in machine learning how dangerous a wrong decision is if facebook misclassify is a face not a big deal right if Amazon tells you how to make thermite slightly bigger deal if an aircraft manufacturer makes an unsafe system and people fall out of the sky that's a huge deal right it would rock the industry trying to be careful so I mean it's a big deal there are a safety critical application so in the in kind of the next part on machine learning and fluid that's that's something we absolutely have to be thinking about is that point one percent cannot be at the cost of all else I will point out we don't trust our models now like our models there's no model for how an aircraft is going to fly except an aircraft line we have uncertainty in every step and there's a whole field that uncertainty quantification about managing the the known issues in our models and so it's not inconceivable that that will be how we handle this in machine learning too so I think we have to be asking ourselves why now you know a lot of a lot of the younger people in this audience kind of you're you grew up professionally in the world of modern machine learning and so you know that's just where you where you come from a lot of us didn't a lot of us saw this massive sea change in how we interact with data and how we build models again I'm gonna oversimplify this is again that image net data set so why now it's partly because of this incredible almost you know Herculean effort of this label training data it's also because our computational hardware is you know incredibly fast and powerful so I don't know you know how many of you track Moore's law and where we are but in my lifetime I've seen the speed of individual cores basically level off because it's hard to make smaller smaller dies and so now we have thousands of cores okay and that has allowed us to and these are not great for you

Channel: von Karman Institute for Fluid Dynamics

Views: 869

Keywords: Big data, machine learing, data-driven, von Karman Institute, Université libre de Bruxelles, VKI, ULB, model order reduction, system identification, flow control, machine vision, pattern recognition, artificial intelligence, mathematics, mathematical tools, discrete LTI systems, Steve Brunton, neural network

Id: wSOClPEsAcs

Length: 36min 6sec (2166 seconds)

Published: Fri May 29 2020