Jul 22, 2020

Don’t know your Supervised from your Unsupervised learning? Do you have your Neural Networks in a not? Fear not, in today's podcast,  Dan and Lee navigate the types of learning that machines can do and how these algorithms work and their application. 

________________________________________

TRANSCRIPT For this episode of The AI in Education Podcast
Series: 3
Episode: 6

This transcript was auto-generated. If you spot any important errors, do feel free to email the podcast hosts for corrections.

Hey, welcome to the AI podcast. How are you, Lee?
I'm good, Dan. I'm ready to rock and roll and go again.
Fantastic. It's a lovely day out there, so let's get started. Um, yes.
So, in the last episode, uh, I tried to step through very simp You did a good job. You did a good job.
Thank you. How machine learning works at a high level to try to demystify some of it. Um I suppose one of the stages though involved choosing algorithm towards the end of that podcast when when you listened back and kind of came up with some com comments on it in my kind of wine based example. I suppose um you know we really started to think well the next logical step would be to look at some of these models and look under the covers at some of these kind of models that could be used when we thinking about machine learning because that's the crux of it really. You've got the process which is the last episode but now you know what are these kind of terms and what are these kind of um uh I suppose models which lie underneath the the machine learning itself. So let's let's clarify from your point of view um what ML is and why it's so different to conventional programming I suppose.
Yeah good place to start Dan and you're right because there is a lot of complexity we're not going to get to the in today's to unpack all of the depths of it because when you start to go digging down on all the different ways that ML machine learning works, it's deep, deep, deep. But it's a good way to think about it that high level because probably most of our listeners at least have some experience of programming in conventional ways today. Whether that be through you know sort of logic languages like logo and other things or whether it's actually in you know proper full you know rich programming languages and you know the the sort of C++es, Javas and JavaScripts and so on. What I guess you get to is this is when we think about programing In the traditional sense today, it's that conventional logic. You get a repeatable and expected outcome. Every single time you do something, you know, you you put a piece of code into the computer and every time it runs, the same answer comes out because there's no data to drive a different differentiated outcome. In machine learning, you've still got the code and it might be written in, you know, could be written in um in Pearl or it could be written in Java or it could be written in um in R or something. Still code, but the code is being fed by data and the data creates this highly probabilistic outcome. It's basically learning what the data tells it and changing the output based on the data. So there is no repeatability to it.
Yeah.
Uh so you know when we think about logical conventional logic 1 plus 1 always equals two. But in a probabilistic machine learning world 1 plus 1 statistically equals two but it's a probabilistic outcome. It probably equals two. And that's that this sort of very unique nuance thing but um but we'll probably dig into that as we get into some of the models
when we look at the the types of models we've got out there. I I hear a lot about supervised and unsupervised learning. What what does that mean in terms of machine learning?
Yeah, it kind of conjures up this picture of uh a person in a white lab coat watching over the computer as it does it job, you know, and it, you know, and it's it sort of is, but isn't that?
Um, so look, yeah, you've got supervised and unsupervised as the two broad domains of machine learning. And it really comes down to this simple idea of um how do we best describe it? It's labeling. So the data that you're giving to the computer in order for it to make a decision is potentially labeled or not. So let's say you have a 100 pictures of cats and 10 pictures of dogs and you feed it to the computer and you say here is a picture of a dog and I've labeled it dog and so you can see that it's a dog. That's ostensibly supervised learning. That's where we are telling the computer what this thing is and saying to it now I've told you what it is. I want you to remember that and learn so the next time you see one of these things you know that it's kind of a dog. Yeah.
Um, you know, and of course that introduces that situation where you go dog, dog, dog, and then you show it a wolf. Well, a wolf looks like a dog. Your human brain and my human brain can immediately tell the difference.
But to a computer learning in a supervised way, it just says, well, everything else I've seen looks like a dog, and that looks a lot like everything else I've seen, so probably is a dog. And what you end up with is this um probabilistic outcome. I say it's 85% likely to be a dog because I can see a smoo few small pixel variances that different but mostly looks at so that's supervised um and and look when you dig into the depth of that you've got all these things around types of supervised machine learning so decision tree so you can kind of imagine what that's like it's a a model that makes a decision in a treeike way and fans it way out based on decision makes things like nearest neighbor so when you start to look at images uh looking at pixel colors and the nearest color of pixel in each part of the image might be considered to be the nearest neighbor as a data point and then you start to tell well if all the colors look like these two things similar to other that's the mark of being the dog um and a whole bunch of stuff like that. Yeah.
So you often see supervised learning in things like image classification like the example we just gave.
Yeah. So when so when we looking at unsupervised learning then so that's complete opposite right?
Well pretty much yeah I mean it's it is but there's obviously a lot of depth of difference behind why we might do the two but unsupervised learning is exactly that where you say okay here is a thousand pictures and I want you to figure out what each one is. You know, you tell me what you see in the picture and I won't supervise you. And that that's kind of where the terminology comes from because it's saying, "I'm not going to feed you the answers." Yeah.
From an educational perspective, you might look at this as, you know, the teacher giving the students a task to do with no context and having them figure out what path to take versus the teacher saying to them, "Here's a body of work and I'd like you to tell me the answer to the question at the end, you know, and here and here's some data to feed you along the way." That's that comparative point between the two.
So, with unsupervised,
what What's really inherently different is we kind of don't know what we're going to find out. You know, that's a an odd peculiar outcome is that in in supervised we're telling it learn about cats and dogs or learn about these things. In unsupervised we're saying here's some information, classify it for us. Tell us what you see and help us look for the needle in the haystack. And that's kind of the the classic interesting problem of an unsupervised model is let's say you give it a computer, you gave it say I don't know a spreadsheet with a million rows in it and there's all these data points on product sales for a big software company over the last 30 years in countries all around the world. You can imagine the data set will be huge and we want to say to it well why do people buy Microsoft Office on Tuesdays?
Well then you look at the data you don't tell it that's what we're looking for so to speak or you don't tell it that's what you're looking for what the data is.
You say show us where the anomalies are. Show us where you see spikes in the data. Show us where data tends to cluster around. around a common set of parameters and that way that's unsupervised because the data the machine is not looking for a thing it's looking for patterns in the thing that make that make sense.
Yeah. Yeah. No absolutely. So clustering is a good good example there to try to and I suppose if you think about there uh a good example as well would be um just based on your examples there you know if we're looking at classification in the supervised way but then you look for anomaly detection then you can't take a photo of every every anomaly that could happen with a pipe for example. Could you so so it's about looking at those images and then finding something that's different maybe.
Yeah. And and look and and you're absolutely right because this the classical problem is in a in a supervised scenario for example if you were trying to do something as binary as this is a cat this is a house you know here's a picture of cat here's a picture of house you could give a computer model with about I don't know 20 30 40 pictures probably
and I would say from there on the computer could accurately tell you that's a house. And that's a cat because there's so many variances in the two. Cat and dog might be different. Cat and you know dog and dog and wolf obviously much less and you'd need a lot more data to do it. Yeah. Because you need to give it lots of data. Whereas in an unsupervised model
the data set in itself determines the scope of the accuracy of the prediction if you like. So the more data you give it the more the computers having to look at more potentials and then make more correlations and then look for more clusters or anomalies. And it's only going to be as good as the data set you give it. So you need lots of data, but but you know the good examples, you know, anomaly detection, medical imaging, you know, we we produce globally millions and millions of medical images of, you know, a number of very serious conditions, you know, and and doctors look at them today and they make an assessment by looking at the image. There's a lot to look at. That's a perfect example where under the right circumstances, put lots of those images together and tell a computer, look for the anomalies. What a computer may actually do is identify things that we can't see because what humans lack often is the ability to see the big picture. You can't look at a thousand pictures and not get fatigue by looking at all these thousand pictures and starting to mold them into one. Whereas a computer looks at a thousand pictures and sees everyone as a unique picture and is able to kind of address it in that way.
Does that mean there's a sort of middle ground here as well then where you've got the supervised learning that you might be focusing on a project but then you need to have unsupervised elements to it. Is there a middle ground?
Yeah, there there is. And it's really unimaginatively titled semi-supervised learning.
There we go.
Because it's absolutely that middle middle ground which is you know some you want you want the power of unsupervised learning where the computer really can help you show things you couldn't see.
Yeah.
But you want the accuracy and again the kind of the logic of processing of supervised learning where you're telling the computer how to follow a particular path. And somewhere in between there is that almost perfect middle ground of of you know an intelligent decision made by human uh injection of knowledge into the process is I guess one way to look at it
and understood. Got it. So supervised learning and then unsupervised learning and then semi-supervised learning. So we're getting we're getting somewhere with these algorithms. So the other thing I hear quite a lot in u the news and in the industry is around reinforcement learning. Now that that sounds like my days when I was at school. Can you explain a little bit more about reinforcement learning?
Yeah, it it sounds like that time when the teacher comes around and clips you on the back of the head because you weren't listen.
Yeah, that's exactly right. Yeah. So, look, but interesting enough, that's a good analogy because it is reinforcement learning is kind of exactly that. And it's a little bit controversial, at least in my mind, because well, if you think about it in the context of a of how you might teach a child to learn and there's a lot there's a lot of similarities between the way in which we teach computers to do things and the way in which we teach children to do things. You know, it's often wrote learning or or example learning. You I do this, that's how I do it. You should copy me. But When you think about that in the context of children and you and I both have children so we know this one you know there's a very different outcome to when you give a reward for something versus when you give a punishment for something. So when we think about reinforcement learning it's exactly that what it it requires uh what they call an agent uh this idea of some third party that's observing the decision the computer is making and then rewarding a good decision or disciplining a bad decision. More often than not it's a reward of a good decision. We don't we don't want to favor bad decision. decision making because that can lead to a preference for bad decision-m. You know, if you like if you if you rewarded your kids every time they were were naughty, they start being naughty all the time. And the same thing is true for computer, it will start seeing that that's a good outcome. So that's the the the thing and it it's another machine learning algorithm and it uses sort of trial and error to churn out the best or the most preferred. So you often see reinforcement learning in uh personalization experiences. A really good example of where we've used a model like this is in the box experience. So if you ever use Xbox and you kind of you get shown content on your home screen and you get fed information about games and events and things that are happening, that information is being fed based on how you interacted with the Xbox service. You know, how long did you look at an advertisement for a driving game might indicate you like driving games. So we'll reward the system to show you more driving games because you have shown a preference as the agent. In that case, you're the agent because you tell the eight the computer what to do. And And that's how you're doing it. So you get this idea of the system learns because you tell it what to what is good and what is bad
and you need to be just conscious of that side.
This is an interesting concept now. So I'm I'm used to the training of the the things like your kids or like your dogs and things like that at home and pets where you're training things for the good sites. You want to you'd assume that you'd be teaching the the machine learning algorithms for the good. You know, you'd be saying, "Yeah, positive result. Great, great, great. But it's interesting you've brought into this now the notion of kind of decision regret, you know, the bad decisions.
Yeah. Yeah. It's a um it it's it's an interesting if you sort of academically speaking, it's an interesting side of how computers and decision modeling can work in machine learning because there is always a flip side as you have a if you teach a computer a good decision. This is what I like. This is what I want more of. You have by definition all Alo taught it what I don't like what I regret you doing for me. So you know you in that we have this notion in machine learning of of decision regret and there's a lot of research going on into this area of how do we how do we actually make that a value point of the decision. So how do we teach computers to make better decisions by helping them under by embibing them with this idea of regret? You know how you and I might do something and go ah I really wish I hadn't done that and we choose to you trigger in our brain this idea that we won't do it again. It's giving computers that sense of doing that but without it being a reward. You know to my point earlier that if you reward bad behaviors you create an outcome where bad behaviors are craved. So it's it's a really you can see these these things become quite almost philosophical in nature even though they're scientific in deployment and and and and operation. It's very you know you sort of introducing these humanistic mindsets.
Yeah. And it's it's very nuanced as well isn't it? Because I think the thing that jumped into my mind you know the things that jump out to me from like you said Xbox, but also like Netflix. You know, I watched something on Netflix last night. It kind of gives you selections based on what you've watched in the past, but I selected something really bad last night. It was a really poor film. And um and you know, you you you got to you got to think about where you record. You know, I didn't record that I didn't like it. So, that algorithm probably still thinks I enjoyed that film, you know, but there again, I did stop it halfway through. So, maybe the algorithm brings that into account, right?
That's a really excellent point. Actually, Netflix is probably even better. example than the Xbox One because it really does play to that. You choose to watch a show and then of course the Netflix algorithm is constantly figuring out well if you like that then you must like this
and then you choose to watch something that you don't like. So you watch five minutes of it and then switch it off. Now there's a data point. The data point is Dan watched it for five minutes and switched off. So we have to teach the machine learning system that that's the that's the moment at which that's a bad that's a regret decision. So we need to tell it to no longer do that and by definition call that the bad decisions and call the other ones the good decisions. And that's a difficult thing because you might watch it for five minutes, decide you don't like it, then go and read a bit more about the show, decide that actually I'm going to give it another go because I think it's getting to somewhere that I might really enjoy, but the system has now decided that's a bad decision for you. So there's a lot. Yeah, there is those nuances. But the Netflix is a really good example because what it introduces is this idea that we always think about data being the thing that you you inherently do that the system will go, "Ah, Dan did that." Often data is also about the things you didn't do. So when you swipe through the carousel on Netflix of all the shows that are trending today, you're actually creating data that says Dan doesn't Dan is not interested in these shows that are trending today.
Yeah.
So we need to change that down.
Yeah. Really? Yeah. It's kind of there. I you know the thing that I think that's a really good example of this because it's um I think a lot of people would know about it is do you remember the um the Google Alph Go engine?
No, I don't remember actually.
So it was um I'm trying to think when was it was probably about oh I was about five four or five years ago they they built a reinforcement learned engine
to do to learn to play the game go which is arguably I'm not
oh yeah I do remember that yeah sorry yeah that's right
and they built a machine to learn how to play go but and this was the you know the amazing thing that popped out of that is the world champion Lisa Doll who's this 18 time world champion played it at so many games as I said probably what is considered to be one of the most complex games on the planet of a human created game. And when uh Deep Mind played him in that game, what he they won, you know, both won a few games, but Deep Mind Alph Go played a move that had never ever been seen before by a human player.
And wrap your brain around that for a second. You think about chess or a game you play and you think to do something that it's been around for years, hundred. It's a very old game. And it the computer thought of a move that had never been played simply by taught by reinforcement learning being taught all the games that humans have. It's fantastic.
Something that we've always done and found something we could never even think of.
It's mindblowing when you think about that's the power of what a m reinforcement learning engine possibly could do.
Yeah, absolutely. So, we've got a structured learning, unstructured learning. Then we looked at some of the reinforcement learning. Now, one of the things that that I remember looking at in university um was neural networks. Uh that was kind of the nearest we got to kind of AI um at that particular point. And it sort of seemed seems a little bit even though the name seems quite fancy or whatever it's quite an old kind of term I suppose and you know like I said it was it was around when I was in university when we talked about our uh the history of AI we talked about neural networks and how that was kind of um one of the ways that people started to look at but do they still fit into the ML story these days?
They absolutely do and I think you just you just tripped over the edge of the rabbit hole and now you're falling falling falling falling down the rabbit hole thing. Yeah.
So So yeah, look, but it's a really important part because it is a it's a it's a a key part of the machine learning framework.
Um, so probably let's break let's break them apart. Let's before we go into some of the history, let's go break them apart. You've got machine learning as we've been talking about which is this sort of broad set of models or algorithms as they're often referred to that we use to discover patterns in data either supervised or unsupervised.
You've got artificial neural networks. So ANN or which is the neural networks that you're talking about which is a sort of a sub set of those machine learning algorithms, but it's used to discover and build the connective patterns in very complex data. And we we'll get more into that and kind of what that means, but it's more about that. The key word is there is the connective tissue side of things. The way in which points connect together
and then you often the other one that you do hear a lot about is deep learning. You know, deep learning neural network is a phrase that you will see banded around quite often.
And deep learning is really more about the closer to the metal in terms of the technology. It's actually more of a the field of of hardware architecture, the silicon and the software integration layers to create these computationally powerful systems, you know, FPGAs, GPUs, and the world of of sort of silicon alliance.
And and I I had I had a call today with a with a education customer actually and we we talked about this and just to bring the podcast listeners up to date for a minute on the podcast um just where where we where we talking about that was that um one of our team was explaining to the educational um team in this educational establishment in Australia about the nuances of the machine learning the the capabilities inside Azure and then actually had a illustration of uh satellite image recognition basically and and how long it took to how many steps it took to train that algorithm and there was three elements through one was the CPU so it showed that it was analyzing you know whatever it was six images a second and then the GPU was on like 50 a second and then the FGPU was on like 5,000 a second. So, and it was a visual representation of when you can, you know, depending on the
not not only does the machine learning algorithm matter, but then the amount and the type of data you'd be doing, you might want to then bring into play those deep learning elements, you know, which bring into play GPUs, FGPUs and and the like, I suppose.
Yeah. Look, absolutely. And there's a and that's really what has been in the last 10 years what's accelerated this capacity of for neural network is the fact that we've been able to model the neural construct in silicon you know in GPUs and NFPGAs where we've created this capability for the logic processes inside the silicon to match the logic processes of the neural network so look and it's probably worth let's go back to the you know what is that neural network piece
yeah
because it's to me this is the thing when when when we talk about And I was looking into the history of this. It's one of those funny things. We started talking about neural networks or the idea that a neural network could exist, a computer system modeled on the human brain back in the 40s, 1940s, not when you were at university.
Yeah. Yeah. I know it was old. Yeah.
And and you know at that point then medicine on the brain, so understand the brain, we were still using sores and chisels to explore the human brain. You know, that's how kind of archaic our thinking was around the human metabolism. It was only in the 70s when we started using MRI. eyes medical resonance imaging to understand the sort of the structure of the brain. So it's kind of crazy that people were thinking 30 years before that that the brain looks like this complex set of systems but this is where the name comes from. So neuronet network comes from the idea of the the the way the brain works which is about neurons and nerve cells and you know we're not doctors but was we understand you know we our brain makes decisions for us through a series of connections and and and what we call synapses. So you think about it if you go to put your hand out and touch a hot fire. The nerves in the end of your fingers, the neurons will tell you will will feel the heat. They'll feel that heat and they'll use sinapses, these messaging between the connections. You know, we talk about the connection being the important thing to connections to tell your brain that, hey, I'm really close to something hot. You know, what are you going to do about it? The brain's going to go, well, you know, I've touched hot things before and it really hurt. I didn't like that at all. That's a learned experience. I'm going to tell you to not do that. So, the brain then sends another nerve signal through the neural network to tell your muscles to contract away from the fire. And that's that is I find to me that's a really that's a really obvious example of how a neural network works.
Yeah. And and it kind of you can see the reason why you know initially when we looking at the history of uh AI and and machine learning in a previous episode, you can see why some of the mathematicians thought that they could model the human brain and and do AI quite simply because it was a case of, you know, copying the human brain.
Yeah. Well, and that's and to think that you were trying to copy something that we really didn't even understand at that time.
But but we recognized the connection that says that something as distinct as my fingers to my brain to my muscles must be connected in some way.
And so what we've obviously learned over time is that there are these, you know, nerves are cells and we understand how now the anatomy of a cell works. And these cells have uh the these edges to them and they have these sinapses that are the mechanism by which they send electrical pulses to other cells to signal, warn or behave in particular ways. And that's kind of what a you imagine a neural network to look like, it is exactly that. But instead of the uh you know there being actual nerve cells and you know back to our quantum episode the other week where we're talking about you know quantum units of uh of the physical world, we're talking about a compute node, a little piece of the compute logic that looks at the data presented to it and then essentially weighs up what it sees in the data. So depending a lot on what we're asking the neural network to do, you know, things like image recognition is a is a really good example where the the neural network's looking at the image from multiple perspectives. What do the pixels say? What's the major what's the big picture? What are the color variances? All of that data. It uses all of that and each time it looks at a a piece of data, it weighs up the pros and cons and then has this waiting process that says, you know, based on what I see, I'm going to weight this more towards parameter X, water more to parameter Y in the same way that your nerves will go, that's hot, that's too hot, and make that kind of waiting process. Uh, and it just does that and but it does it then of course at a exponentially bigger model than, you know, our human our ability to use our human brains can manage because it might look at a million data points in a second.
Yeah. And and that's the that's the thing that that struck me in the last episode when I was just comparing two variables, you know, it it makes sense, but then you add a third variable, a fourth variable, multiple data points, it it becomes uh quite complicated very very quickly.
Yeah. And look, it's it's not nearly as bad as the quantum story, but it's still quite a hard thing to get your head around is that you're looking at the the multiple layers of information that are in the picture or in the data to make a decision. And and the funny thing is, I mean, this is this is the the the nuances of the human brain is that You and I, my a child of five can look at a picture and immediately go cat, dog, horse, duck, whatever it is. A computer has to be taught that. And a neural network is a really good way of doing it. But you see the computational complexity to actually calculate something as simple as the differences between images. And in our human brain, we could do it after seeing a couple of pictures. You could show your kids, like yours, because they're much older now. You know, we could show our kids, young kids, pictures of ducks and they pretty quickly figure out that you know uh a condor is not a duck because it just doesn't look the same. But we see that so quickly and it's that synapse firing that our brain just making so many connections
that a neural network and and you know that this introduces this whole thing of to solve a very small problem in computing like you know image recognition requires massive amounts of compute massive amounts of compute and compute costs energy costs uh you know impact to to the climate when you add all these connections up around the cost of energy and the power the usage efficiencies and all that and there's a whole thread that we maybe will talk about in one of the future episodes around this red versus green AI you know which AI is good and which AI is bad for the sort of the for the planet but anyway interesting point
it it is and so just so just to extrapolate a little bit more and just to fill gaps in my knowledge as well so if if you've done a machine learning algorithm and and the machine has learned uh a particular problem to solve and then you've got a can can a neural network then be used. Can you then extrapolate that into a neural network to be almost like the the brain of a um uh a robot or whatever or a machine to kind of connect multiple um algorithms together or is it not like that? sort of not quite like that. I mean, that's an interesting place of where we want to get. So, what you what you're really hedging towards down there is the difference between general AI and narrow AI. You know, today we take data, we might use neural networks, we might use basian trees, decision trees, we might use all different types of models to train an AI model to be able to infer an outcome based on what it's been taught before. And you can take all of that and let's say it's about uh recognizing you know, trees and water in image data, that kind of thing. You can teach it all that and then we can put all of that into a tiny little compute, something like the um is it the Nvidia Jensen chips? They're about a credit card size AI chip. Put it all that put it all into a drone, have the drone fly over an area, and it'll be able to tell you the difference between forests and trees. However,
it's narrow. It only knows how to do that thing. What you're asking about is this ability for the truly the the connection tissue between what I've just what I know what I'm learning as I go on and to be able to connect that to all of these other data points to go. I'm seeing trees, I'm seeing water, and now I'm seeing a building and I know what buildings are. And because that building's there, it's probably a boat shed because it's next to the water. And you know, you can see how the picture fills out
and you're heading towards the singularity if you want to get into that world, which is, you know, when a where a computer system is is able to
behave like a human brain in that way. And that's neural networks connecting to neural network. works with all the data you can imagine. So, I it's where I think it's an interesting future view, but we're not quite there yet. I think it's fair to say.
Good. Interesting to see how it all fits together, though, because I suppose it's tied together a couple of the last episodes really because it's tied together the fact that we've got these algorithms, we've got large supercompute and going into quantum, you know, the fact that this data can be processes very, very quickly. So, it's quite easy to put all these episodes together and see where we can be in four, five years time uh and how the impact of AI is going to actually become even more important to the way we live our lives I suppose and how the environment gets impacted and all kinds of other areas. So, so it's really great I suppose to look at ML and that's given our listen listeners a real solid understanding hopefully in the last couple of episodes of what it is and all the algorithms associated with it. In the next couple of episodes, let's think about speaking to people who might be able to bring this to life, I suppose. What do you think?
I like I think, you know, it's great. Uh who wants to hear a couple of old crusty men talk about something when you want to hear the kids, you know, the people that are going to be living this life and building these things. Let's let's get let's get their input.
Yeah, definitely. Cool. Well, thanks again for sharing your knowledge today, Lee. I really appreciate it. It was another mindblowing episode and hopefully everybody's listening in their cars are kind of safe and pulled into the sides and they can carry on with their journey now.
Love it. Yeah. No, Dan's a pleasure. I always enjoy talking to you, mate.
Thanks so much. That's brilliant. Cheers.
Thank you.