Jul 8, 2020

In this episode Dan talks through the machine learning process.  What steps do we need? What data do we collect?  And why does thinking about alcohol make this easier?

________________________________________

TRANSCRIPT For this episode of The AI in Education Podcast
Series: 3
Episode: 5

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

Hi, welcome to the AI and education podcast. I'm Dan and Lee's going to join us a little bit later to add some of his flavor to machine learning and give us some more insights into how this works. But what I thought we'd do today's episode is actually look at machine learning first of all and think about the steps we need to create create a machine learning model. So this is kind of looking I suppose under the hood at machine learning. Let's start at the basics. I suppose the world is is filled with lots of data. Um whether that's music data, video data, documents, spreadsheets, there's data everywhere. There's data collected from IoT and all sorts of things. What machine learning brings to us is the promise of analyzing that data and getting meaning from it. Because the more data that's been collected, And it's being collected everywhere nowadays from our cars to our refrigerators to our fitness trackers on our uh watches to our mobile phones. There's data coming in from a lot of different points. So the more data that comes on, the more difficult it is to actually manually process that data. What we've had to do over time, I suppose, is actually manually rewrite rules to adapt systems based on the data that we can process ourselves. But the more more and more data becomes abundant, it makes it more difficult to manually create those rules. So, think about machine learning as as not being a dark art and this thing you need a maths degree for and you need to go into major major detail. There are machine learning and data scientists out there, but all it is is a set of tools and technologies that you can utilize in your organizations to answer questions with the data. There's a lot of debt data like we said being generated not only by those people but devices and machines and this is going to continue to grow uh and like I said we've manually written rules to adapt to these systems at the minute and the volume of this data is surpassing our um ability to manage this data really well. So we've got to find a way to manage this now automatically. So let's think of some examples. So image tagging say if you're using some social media platforms like Facebook, image tagging is is available in in platforms like that. And it's interesting because a lot of that is automated these days. So there's a lot of automatic tagging of tagging of images that happens and that's using machine learning. Um things like Netflix or Spotify that recommend playlists to us and recommend next videos to watch. That's all machine learning algorithms in the back end knowing what our preferences are. If we prefer action movies, for example, and sharing with us what other people think are good and and sharing those um insights with us to recommend us to watch the next movie or listen to the next audio track or band. So search is another good example using Bing or Google and line those searches are machine learning algorithms and they are looking at the text of what you're searching for and also they adjust your results obviously based on the interests that you have where you are, your location, and really trying to think about what you entered and giving you good results there. So, for example, if you were searching for something specific around um Java is a good example. Are you looking for a coffee or you're looking for a programming language or you're looking for uh a country to go on holidays? So, you know, they adjust those search and optimize those search interests based on machine learning. There's also lots of things is in fraud detection and this can apply to things across multiple industries including health care. Say for example for skin cancer detection, self-p parking on your cars. There's a heap of uses of machine learning. So what's the process for this? So let's look at this example. Say if we're looking for a model to set up to tell the difference between beer or wine, we can use machine learning to help us with that. So if we are going to use machine learning for these we need to test some variables obviously. So if we've got beer and wine um we got to think about what variables we could test and there's multiple variables we could test here. So let's just use this example at the minute um say alcohol content and then the actual color of it. Um how would you record those? So color would be obviously through wavelength of light and alcohol percentage would then probably be through here um using a hydrometer to find out how strong a drink is. So you know how strong a drink is and what color it is, then you should have two good indicators of what that uh drink would be. There are multiple uh variables we could use, but actually for this, we'll just use those two features of the drinks. Okay, let's look at the steps of machine learning here. Step one, we get the data from the bottle store. We've got the bottles we've of wine and beer. We'll then use the spectrometer to collect the waveform data um to get the color And then we'd get the hydrometer to work out the alcohol percentage of each one and we record those in a table. Nice and easy. Step two is to prepare that data. So we don't want the order in which we collected these things. Say if we did all the beers first, we don't want that to mess up our uh training. So we'd actually randomize and prepare our data. We want to determine that uh it's random and then we can create some visualizations to help see any relationships. So for example, if we've collected more wine than beer, then we want to know that because we don't want any bias to appear in our training model. So once we've actually collected that data, we then split it into two parts and it's usually like a 7030 rule or an 8020 rule where your training data might be 70% and 30% would be actually test and evaluation data. You'd want to keep a clean set of data to test and evaluate your model at the end. Uh you don't want to use training data for that because it already knows the answer to that data. So we'd put that test data aside and save that for later. So now we've got our set of uh training data that we can use all prepared randomly. Next thing we do is choose a model. Now there's lots of models out there that are already created by Microsoft and others for different types of um data. So there's models out there which are good for sequences. There's uh models out that are good for text. There's models out there that are good for music and video and images. But for this we can use a simple data model to compare two variables. Okay, so we pick our model and there's a heap out there that you can use. Okay, step four. Now we've chosen the model. Now we've got to get the training part done. I'll try to explain that quite simply, but essentially imagine you've got a X and Y axis on a graph and you can now start to plot that data. So on the X- axis, say for example, you've got the alcohol percentage and on the Y-axis you've got the color in wavelength. Um you can sort of plot the different front uh sample data and it starts to scatter across your graph. What the training data does is try to separate that out um by putting a sloped line in and the training data will go for each step and try to move that line to get a best fit so that you know what type of uh uh drink is on either side of the line. Which ones are beer and which ones are wine. So just that slow of that line using the formula we all learned in school y = mx plus b. It would uh adjust the slope of the line to best fit. So you'd run through multiple steps. You know the first step might be wildly inaccurate. It's just a random line through the beginning of the the data right in the middle for example and it'll adjust that until it gets more and more accurate and has a really good representation of all of the wines on one side of the line and the beers on the other side of the line. So that that inter line the slope will adjust in each training step. So you get a really good accurate model. So we know then when we plot the next um piece of data if that data is on uh the upper part of the line it'll be beer for example and the lower part of the line it'll be wine. So hopefully you can visualize that. It's hard in a podcast I know but the idea being that you put the data you scattered it across and then you start to add your model. In this case it's going to be a simple formula and a slope which is going to separate two sets of data between beer and wine and that's adjusted in training steps to go and and allow you to put data in. So step five is when you evaluate this. So we grab the data we used earlier this the data we set aside and we put that in and see if the answers that come out are correct. We test it against the data that hasn't been used for the training purposes and that's supposed to I suppose represent the data in real life. Uh and again like I said you might 8020 depending on the size of the data you've got. If you've got a lot of data, you might only need a small amount of training data, say 80% to 20% for training. But if you've got a say a smaller data set, you might use 7030 just to get more accurate training data to test your model. So once you've done that, you know that this is pretty good, then your data scientist come into it or depending on the algorithms and models you're using, you can do what's called parameter tuning. Now I don't know a lot about this, but um essenti what you're going in you can tune other variables. So for example you know the data sets that we've got in for the the beers and wines that we've collected for example we haven't included some of the variables like zero what happens if something comes in where it is purely black or purely white or water you know what what about some of these outliers so it's thinking about tuning some of these parameters so that you can actually really make that model more and more accurate. Step seven is getting used to actually answering some questions with these. So you this is when you realize the value of all these steps we've done so far. So is this drink wine or beer? And we can determine that by capturing the wavelength and the alcohol percentage of a particular drink and putting that into our model. So that's the fun starts and you can actually use that on real real data. So let's just go through those seven steps again. You gather the data in you then prepare that data. Make sure that it's in the form format that um you need. Make sure that the data is randomized and separated out between training data and your um actual test data. Then you choose a model depending on the type of data that you are trying to put machine in again. So that's images or audio or text or numbers. Then you actually go through a training process and you use um your training algorithms and go through training steps to actually train that model. It gets more and more accurate a bit like was driving a car. When you start to learn to drive, you know, the more real world data you get in, the more times you reverse park, the better you get at it. So, it learns and learns and learns and trains and becomes more and more accurate. After you've done that, you then use your evaluation data which you took from the preparation stage to actually test with some uh independent data uh against your algorithm. Then you can do some hyperparameter training which is tweaking the edges of some of these variables and then you can do some real world predictions and some modeling. So hopefully that's taken us under the hood of what machine learning is. You know, it isn't uh dark art. It's actually tools and technologies that we can use to make sense of data automatically when we get in lots and lots of data in. So I really hope that's helped and I'm going to bring Lee back into the studio and get his thoughts on some of the machine learning concepts and some real world examples. So, what do you think, Lee?
Hey, Dan, it was great. Great uh great story. Great way to explain it. I I got to ask why you went to alcohol as your first point of preference for a model to work out the two things.
Maybe that says more about you than anything else.
Yeah, so it's that it's that process, isn't it, that that I think when we think about machine learning, it seems like a dark art, but actually um it's a logical process of steps that people have got to go through.
Yeah. Look, that was what when I listened, you know, to kind of where you took the story and and that was the bit that struck out for me is it it is it's a bit mystical. We kind of think machine learning it's this blackbox thing that computers do that nobody else understands and there is some depth to it and complexity to it for sure. But reality is when you break it down into those seven steps as you did and create that that journey from getting data to sort of evaluating the data to making a decision based on the data. It's pretty simple in many ways. Um you know, and and and a great example of how we can, you know, when you walk through those seven steps you've got that anyone, you know, in education or in in in any domain could quickly kind of start to learn a bit about this process. And that that's what we want to get out of this, isn't it?
Yeah. And interestingly, I suppose you just alluded to it, and I haven't mentioned this at all yet, but you said right right in the middle of that sentence that it was about how they use that data at the end or those results of the the machine learning. And I suppose being able to go back and justify the decisions you've made. If it's telling me that, you know, a student or a particular type of oil or whatever it might be has an impurity in it uh due to machine learning, then you have to be able to step back through because you might make a big million-doll decision or a decision based on a student's life or somebody's health on that piece of machine learning. So,
yeah,
that's a really interesting point as well, being able to go back and kind of say, well, this is how I came up with that. decision, right?
Well, and absolutely. And it's interesting when you think about the seven steps that you talked about, and I'll refer back to the middle of that block, which is I think it's step two, data preparation, step three, choose the model, and then step four, training. If you think about what that model actually is, the preparation of the data, choosing a model, and training, that's almost should be a cyclical process because that's the bit where you you you introduce the data to the problem. And and you know, data, as we've talked about, I guess in other conversations, you data has these biases and has these challenges. you think about and then the model and I think maybe next next maybe maybe next week we need to really unpack that model piece and think about what is the the way that models influence the way that data is used to create the outcome because that's the training process you know a model doing it and being trained in one way and the data that you feed it creates the outcome and the evaluation as it were that you that you link to so I think that you know that's where you guess there's a bit to be unpacked there isn't it in that that kind of complex middle bit
yeah yeah definitely definitely so let's catch up in the next podcast episode then and really try to extrapolate this a bit further. What do you think?
I'll I'll pull all that other quantum stuff from the one before out of my head and let's fill it full of machine learning stuff. Let's do it, Dan. Look forward to it.
Fantastic. Cheersy.
Thanks, Dan.