Models and Interpretations

By Stephen Downes
Dec 09, 2021
Transcript of Models and Interpretations

Unedited audio transcription

Hi everyone. I'm Steven Downs, welcome back to Ethics Analytics and the duty of care, We're still in module seven, which is the module on the decisions we make. And this involves looking at the full artificial intelligence and analytics workflow and looking at all the issues that come up on a day-to-day basis.

As we apply these advanced technologies to practical applications such as learning and development. This video is called models and interpretations It follows the the video on AI, algorithms and tools. And it's the next step in the process. After we've used the algorithm or the tool, the result is a model and then we take that model and we apply it in some real world application.

Now, I want to talk about these models and I want to talk, but how we interpret these models. So we're going to sense of some of the issues that are involved with them. So, to begin, let's take a look at this little model. This is a model of how many days there are left until Christmas.

So when we get to December 25, there will be zero pebbles left. And each day, we'll remove one pebble. So, looking at that, what can we say? Well, that's count the pebbles. One, two, three, four, five, six, seven, eight, nine, ten, eleven. It's that tells us that when I'm looking at this model is December 13.

And there are 11 days to go before Christmas. Now, in fact, it's December 9, but let's not let that distract us. This is the picture I had, but that's the third thing with models. Okay? So, basic concept here, the pebbles are on the model and then what they stand for.

In this case the number of days left until Christmas that the interpretation of the model. And then if I actually use this model and actually take a pebble off the pile each day, that's an application of the model. So that in essence is what I'm talking about. I'm going to spend, I don't know maybe an hour talking about that.

Yeah, pebbles on our table. All right? So how can this get complicated in a hurry? Well let's think about opinion polls. You know, elections. So we know what an opinion poll is basically to survey done usually by a research firm or you know, it might be done. Using people selecting donuts or guinea pigs, walking through holes, in the wall, whatever to collection of preferences for an upcoming election and you've probably seen them right?

The Paul say, the liberals will win, you know, they'll get 45% of the vote. This poll is accurate to plus or minus 5 points. 19 times out of 20, you almost certainly heard. Allah, I've noticed in recent years, they've stopped saying the 19 times out of 20 bit, but they used to always say that.

Well, what does that mean? Exactly, certainly, it's a bit fuzzy because there's a margin error in it, right? We've been told plus or minus 5 points, 19 times out of 20, I just statistical thing. And that's based on the mathematical relationship between the number of people that they've sampled and the total number of people that will vote, of course, some of those things are a little questionable.

How do we know how many people will vote? Well, usually, that's a poll answer, right? Is the number of likely voters, it could be the total number of actual voters or permissible voters, or in the US registered voters. There's a little bit of fuzziness there even more to the point.

What should we think of a poll? Should we take it as nothing more than a snapshot of what people think right now? Or is it useful for making predictions about what will happen in a week when the election actually happens? These are the sorts of questions that come up when we do things, like sampling and statistics.

And I said before, I've said it several times AI in machine learning are essentially nothing more than statistics. Applied statistics, really big statistics, but still so all the questions are. I've just asked a little poll. These are also going to apply to AI in analytics, aren't they?

Modeling can be tricky and there's all kinds of ways we can get things right election. Polls are great because they're in this very limited universe, where there's really only five possible outcomes. Right. That's the number of political parties involved in the race. But look at this chart which came from XKCD.

Wonderful, cartoon created by Randall Cunningham and imagine showing this chart to somebody in 2019 and the lines respectively, stand for the popularity or the demand for sewing machines, webcams and through Cuomo which is just died off flower and pangolin, which I don't know what that is. So we won't worry about that from that.

Something now suppose here in 2019, try to imagine what you think happened in 2020. Now, of course, we know after the fact what happened, there was a pandemic and everybody raced out to buy stuff to make their own bread and make their own clothes. They bought webcam so that they could be online.

We don't know why Andrew Cuomo, got popular. All right, he's the CNN guy. Right? And pangolin. I still don't know what that is. So even if we have this data, they wouldn't really tell us what this data is reporting on. That's a bit of a trick.

What happens when we choose the wrong model? This is a short article by Nicholas T Young. These making the point that he chose the wrong algorithm to predict which instructors teach programming. But what he asks, if he had instead been creating a model to predict what patients should receive extra care and I can extrapolate on that.

What about if we use the wrong model to grade student assignments, or recommend learning resources or learning paths or to predict whether they will fail or to create learning resources for them? This could be a significant problem. So using the wrong model is going to be an issue in the application of AI analytics and learning and development.

So models in AI and analytics are really the subject of what's come to be called the black box question, right? And we've seen this talked about and quite a few articles and and papers people ask, you know, or they say the use of black box models and I'm quoting here.

It makes it difficult for us to determine why decisions are being made paradoxically traditional limitations on accessing data on projected groups can hinder the ability to assess models properly. Well yeah, they would say that. But the reason a problem here, right? We feed data in stuff. Happens and outputs.

The result. Now, we know how that happens, because we just did module on algorithms and tools. So we feed the data in, it's run through various neural network, algorithms that learn or our trained, and then it pops the data, the other end we go through all that process. I don't need to repeat it.

Now won't repeat it because I don't want to, but that's a problem from the perspective of somebody using this, right? We've got a fully trained AI and they're saying, well, I don't know how it's making decisions. It's just the black box. You know data comes in as we see here through the back.

Black box outputs isn't saying nice graph. But what do we even mean by this? We know everything that's in neural network or we know, you know in principle anyways for any given neural network, we could get a complete statement of everyone in every zero in this entire network, no problem.

And and sometimes people actually do do these things by hand so that they can see how each calculation is working itself out. So we can see we can't see the results but there is a sense to the concept of black box model that perplexing. We don't actually seem to know.

Do we? What's going on? Well, let's take a step back, let's take a step way back and think about the concepts behind the question that's being asked. When people ask about the black box and we're all go, is to what all call here. The Cartesian revolution named after Rene Descartes, although you could even call it, the Copernican revolution as Thomas, Kuhn did.

And basically it's the idea that instead of the world being made of analog unchanging, essences that flowed and had waves and impulses, and all of that, we could think of the world is divided into a whole bunch of parts and look at the relations between those parts. Basically, look at the world in terms of quantities or in terms of mathematics and that's what all of these guys were doing.

They were inventing and deploying mathematics, which was new to them. It came via the, the fall of the Byzantine Empire and see influence of Arabic numerals, Western Europe and the realization that they could use this new science algebra from from the Arab Empires from the caliphates to measure things in the world.

And that was much better because as Rams said, you know, paraphrased all the things that Aristotle has said are inconsistent because they are poorly systematized and can be called to mind only by the use of arbitrary mnemonic devices. So what all of these people did over the course of and maybe a hundred and fifty years, 200 years, give or take is to almost rewire our understanding of the world.

Changing it from continuous substances to relations among discrete parts. That would be described mathematically. So there's a good intuitive sense of what was going on here, right? The mathematics that we use would represent or describe or help us predict properties in the world. For example, Kepler, identifying the orbits of the planets or Isaac Newton, describing the motion of objects, all using this new kind of computational device called mathematics well mathematics and people think, you know, there's just one mathematics two plus two equals four that's it.

That's the end of the story, but when we do mathematics, or when we look at the philosophy of mathematics, there are different ways we can think of what it is that we're doing. I've thrown a few meta theories. If you will or philosophies of mathematics, we could think of mathematics as being properties of ideal forms that are actually out there in the world somewhere.

And we're discovering and describing them or mathematics might be purely. The formal relations between numbers or mathematics might be our intuitions about quantities and relations and and brewers interesting. There's there's a thing called the excluded middle and brewer says, note there is no excluded middle and that's like saying,

Not not P is not the same as saying P. So and there's a sense to that if you think about it, you know, suppose somebody said to you you're a criminal and you want to soften that of being you say well I'm not not a criminal, but that feels somehow different of a logically should feel the same anti-realism.

Michael Dammit. These math medical objects that were describing are real. They may exist in our minds, they may exist as purely formal logic, but they're not actually existing properties in the world. And again, this kind of makes sense because mathematics includes the concept of infinity. But in the world, there's nothing infinite.

Is there John Stewart Mill and Phillip Kitcher advocate? A kind of operationalism and what that means is mathematics is just a notation that we use to describe actual operations, like, using stones to stand for sheep and then manipulating those stones in order to add or divide comps of sheep Rickenstein advances, what might be called?

Has something like conventionalism and analogous to the dictum that meaning is use what mathematics stands for is what we actually do. And then there's the concept of computation advanced by I'll entering. And, you know, the the properties of, for example, decideability and completeness. These are very different from each other.

Which one is the right way to talk of mathematics? Well, arguably. None of them are arguably. They all are arguably. Any given one is I kind of fall into some kind of version of operationalism and conventionalism. I'm certainly not a mathematical realist. I think it's very possible. We could have come up with something different to describe the world.

No idea. What that could be though when it's kind of hard to make it so that we can just say, well, when I say two plus two equals four knew what I really mean is. And and we don't know. We don't know how to answer that question. You want to say, okay?

Would like to carry philosophy stuff. Sure. But it comes down to the question of how we prove things because you know we can't just make stuff up, not even in artificial intelligence and analytics. So it needs to be some understanding of how we would prove the conclusions that we draw from all those ones in zeros that were flinging around with these systems and corresponding to the different philosophies of mathematics there are different approaches to proofs.

And this is a bit of a caricature here, but it'll do I borrowed it from James Robert Brown. Who's last name? I forgot to capitalize. Sorry, James Robert Brown. But here's a list, it could be formal proofs. That is to say, symbolic derivations, using symbols, and rules or accidents or the way we prove things might be based on the intuition, just think about two plus two equals four, right?

All you have to do is think about it. What else could it be intuitively? You know that two plus two equals four or it might be inductively, right. The way we can say that say all crows are black or perhaps blackish, because we've seen so many crows, and that's what's always happened.

Every time we've added two plus two, it's come out of four. So probably will. Next time there's also the what they call the hypothetical deductive model, that's the traditional picture of science that so many people have where we have a theory, maybe it's a mathematical axiom of some sort and we want to prove that theory so we use that theory to make a prediction and then we go out into the world and see if that prediction came.

True and that confirms our theory, another way we might do this with pictures pictures can convey truths. You know, the principles of Euclidean geometry were shown using pictures. I can show you a picture of a square and ask you to say, how many sides are there and you say four and how do you know, we'll look at the picture.

There's also diagonalization, which was the mechanism girdle used to come up with his incompleteness theorem or they're even the thought experiment or you have somebody like Einstein asking, suppose I was sitting on a bus that was moving at the speed of light and it encountered another bus coming in the opposite direction which is also moving at the speed of light.

How fast are those two buses approaching each other? Well, there's no speed greater than the speed of light. So, there must be approaching each other at the speed of light and usually go. Huh, think that one's true? I am starting did that's how we got relatively in a lot of the stuff that we have today.

What we do generally, certainly in the realm of theory in mathematics and formal reasoning, etc. Is we use a model and basically the model is going to be the thing that we test our theories against. Here's a simple model. It's a state space. And in this case, the state space is the model representing all possible outcomes for a pair of dice.

If you take a pair of dice, you roll them. And I've been gesturing all along. You've haven't been seeing it, take a pair of dice, you roll them, and then you see what the two dice are. And like, if you're all two ones, there's one there's one, you got a two draw, two twos.

You got a four. If you roll a two and a one, you got a three or if the other dice lands on one in the first one lines, onto you got a three and so on. So what's the probability of getting any value? Well, let's count. How many squares we have?

Oh, we have 36 squares. So let's look at the value, the value of 2, so that's one chance in 36. That's the probability, same with 12, right? But look at seven one, two, three, four, five, six. So that's six chances in 36 or one in six probability that will get a seven when we roll two days so we don't really need to know anything about the actual dice in order to be able to talk about what we're doing when we make a prediction using a model.

This model tells us that think about it. It explains everything right. It's a complete description of the world, if the world consists of two dice and this is all the possibilities. And then everything, we know about the world, what has happened, what will happen in this world can be deduced from this state space and in fact, you know, we we could have a lot of fun.

We, we could even say, well let's take this state space and use it, you know. Extend it back into infinity to represent all the times, anyone has ever thrown the dice and what should happen is in the actual world, we should get a seventh six times. Out of every 36 times, we throw the dice.

We should get a two only once every 36 times. So you're going to four three times out of every 36 times or one in 12, right? But what if the guys are loaded then the state space would not be a good model to use. Would it be the frequency?

Interpretation would probably be a better one. The one where we look at all the possible states of the dice in the past. And we saw that with these dice, at least we were getting rather more twos and rather few were 12s and that would tell us. Okay, well, you know the maybe the the dice are loaded to favor low numbers and to unfavor high numbers.

I don't know how you do that with dice but let's say, so we'd forget about the state space. We'd use the frequency interpretation. But what if we don't know it, that frequency is. Yeah. I mean who's going to see every role of every possible device? Right? It all happened in the past, you're sitting here with dice now.

So you look around you at all these gamblers and they're betting on the outcome of these of these dice and you see some of them, the ones that are going better actually are wagering more on the low outcomes and less on the high outcomes. We don't know why they just are.

Well, that tells us what people would bet on, and maybe that's how we should interpret our model for the possible outcomes of dice. So, I've given you three pictures here, I've given you state space, I've given you a three frequency interpretation and I've given you a subjective interpretation. Of course, bonding to the works of Rudolph Carnet.

One of my favorite philosophers Hounds Reichenbach and Frank Ramsey as different ways of interpreting the probability calculus. So even though we have math, even though we have a mathematical description of everything, there are still different ways of understanding that mathematical description of everything. Well, that's how formal semantics works.

There is a world out there, so we assume. But in certain important respects were world, is inaccessible to us big chunks of it. Lie in the past big chunks of it. Lie in the future. Big chunks of inner outside. Our range of perception. But we still want to be able to talk about it to say true things about it.

But what happened about, what will happen what's in it, what should be in it, what is the case? What ought to be the case. So we set up the formal semantics or we have the model which in this case is a set of statements daisy is a cow cow.

Is a kind of animal. Mary is a person. Person is a kind of animal Z123, ABC is a car, Mary drives Z123 ABC and then that interpretation is in this case, this picture looks a lot like a Venn diagram and in fact, that's what Venn diagrams are used for.

So to provide an interpretation of predicate calculus and we can see right here, are the couch. There's daisy, there's another cow. And here's the people, there's Mary and there's couple other people and containing all of those. Are you see the set of animals and here are some other animals that are not persons or not cows and outside to set of animals.

We have things that are cars. Here's Mary's car, here's another car. We got actually draw a line between to the car if we wanted. And this is our what they call a universe of discourse as we have ordered pairs, a B, which are members of the set this and this formal semantics.

And it's by examining this formal semantics that we can prove that statements are inferences made. Using our model are, shall we say truth preserving or have some other semantical property? So the thinking is this is kind of what happens in the human brain. Roughly very roughly in the human brain.

We've got one of these and one of these and they work together. And so we have what are called. Representations me identification of individuals and how they're related to the rest of the world and our their own networks. Our neural networks contain these, right. That's kind of interesting because there are studies it sure that these mirror representations across species are different from species to species.

For example, a rat. Think it's a right. Like the most rodents. They don't really distinguish between these things per se, they're going to show

They're going to show a representation that includes the cheese, the table, the window, always together humans. On the other hand are going to see the cheese as something separate from the other things. Something that has a continual existence in its own, right? And the monkey is kind of mixed.

It might be able to separate the table from the cheese, but it can't separate the cheese from the room. The cheese is in so different ways of representing. So this is kind of interesting because they're all neural networks. They need to those three brains, but the neural networks actually represent the cheese on the table differently.

Semantic web which we've heard a lot about, is I attempt to instantiate this and computer logic, the logic of the internet and we're doing the same sort of thing recreating models. Where, in this case, the model is an antenna, sorry and ontology. And it has entities or more accurately types of entities like animals birds fish, canaries, penguins etc, and then ways, for these things to be related.

So Paul Schuster, for example, was born in Dresden dressed in as a place. Paul is a person. And so that's how the semantic web creates the interpretation that our models can use. Well in artificial intelligence. The model is a program that has been trained on a set of data to recognize certain types of patterns and we see some examples moving about to the right of us here.

They're not nice neat collections of words or things like that. The way the semantic web is. And in fact, if you think about it, the semantic web is kind of like Rudolph clarinets state space. It's describing all possible states of affairs in the world through series of subjects and predicates and sentences.

Paul is a human, Paul was born in Dresden dressed in as a place. You see in the nice symbolic representations there, that's the semantic web. If we go back to current app, go back to the very beginning clarinet. Basically, your state space is a list of all of the possible entities and a list of all the possible properties.

Those entities can have the dice game. The entities there are six entities, which we will call one, two, three, four, five or which we will call dice, number one, dice number two, and the six properties that each of those dice can have. So dice one, one dice one, two, etc.

So that's the state space expressed as a formalism in artificial intelligence. At least in terms of neural networks. We're not working with words, we're working with ones and zeros. Oh sure, those ones and zeros may have labels, but the labels are separate, that's the human interpretation. That we place are all the ones in zeros that are going on in an artificial intelligence.

So in near on that works specifically we say that a model is the set of connection weights in a trains network and just as an aside that's why I say stuff like knowledge is to set up connections in a network. Right. Our knowledge of the world is some total.

This model that we have of the world except without the world, right? Because the world lies outside our mind and that's kind of the tricky part, right? We develop this model in our brain or a neural network train to neural network that we're using in some analytics application develops.

And, you know, some sort of set of connections, set of weights and values for each of the nodes in that neural network and we so all yeah, let's find what does that mean? What does that stand for? I it goes back to that original question. All right, we have the ones in zeros but maybe they represent the fact that the pandemic hit in 2020 or maybe they represent something else, we don't necessarily know.

You can't tell just by looking at the ones in the zeros and that's the problem.

So we need interpretation of AI models. For example, when we say your credit card was declined or denied, right? We are looking for some kind of picture up a lot like that picture generally that explains or shows why the credit card was approved or denied and it might be a straight line through a set of possible data.

Or it might be a really curvy line through that data. So and it might relate to the late fee and the annual income. But now you have to ask. Why did you choose this? Curvy line instead of the straight line and we might sit down y'all. If we had only those two variables we might say something like well look, somebody's annual income is good.

Were not going to penalize them with late fees so much, unless they go over a certain amount and their income is really good, and we know that they can afford it, but if they're income is so good that it doesn't really matter. Then again, we'll just charge some whatever they want.

And, you know, rich people can be in debt but to a certain point and then we're just not gonna let them have any cards. That's how it's really convoluted, but that was my best guess and interpreting that squiggly line. But again, the words aren't in the data, and that's the tricky part.

It is what we would call subsymbolic. The relations are relations between words except in, you know, labeled AI models, which are supervised. And then, we could sort of say that the individual neuron stand for words along, you know, probably not, you know, they, they stand for. Well, we don't know what they stand for.

And so that's why we can say something or like as Boyden Crawford say, and the text, I referenced earlier in the course, do numbers speak for themselves. The answer is no, and yeah, clearly the answer is no, but look at their explanation for this. They say significantly Anderson's sweeping dismissal of all other theories and disciplines is it?

Tell it reveals an error again, under current in many big data debates, where other forms of analysis are too easily, sidelined, and sick there any sort of gonna go. What? Because that's not what it means at. All right, both Anderson and boarding. Crawford are going to agree that the numbers.

Do not speak for themselves but Anderson's not going to say we can tell the same story, just with words, whereas it appears that Boyd and Crawford want to say, we can tell the same story just with words and suggesting that you can't is. I don't know, an arrogant undercurrent of many big data debates right there.

Interpreting Anderson's response as a kind of arrogance. But the model doesn't support that really. The problem is that explanations that use concepts like the semantic web are just too coarse. There are two blunt and instruments to represent what's happening in inside the neural network. So if we're going to use words, the best we can do is label the input and label the output.

But, you know, any labeling of the entry media steps would be guesswork on our part. We could sort of pull it off, especially, you know, if we built to feature extractor from scratch, we could actually label every single one of those neurons. But then someone like Anderson is saying, well, what does that buy us?

What do we get out of doing that? Have we made it more explainable? Well, no, all the calculations are exactly what they were before. It's just we've forced this interpretation on it.

You know the model that we have as humans can be, you know, it grows it. Develops with experience with practice with teaching etc, with exposure, to the entire culture. Maybe some particular parts of it and we can think of that model as the development of what might be called, perceptual expertise over time or as a summarized, in this review, by John events and enhanced capacity for perceptual recognition or discrimination, with respect to some feature or category.

You know, somebody for example, could be a perceptually acknowledged, recognizer, or discriminator of bird species and just getting really good at that or cars or tumors depicted an x-rays. And similarly and a mod on AI model. Trained on a specific task like bird recognition, car recognition tumor detection. These are also models with a particular type of perceptual expertise.

The question though is is perceptual expertise always virtuous and giving an example gadar. If you're wondering what gauge are is, it's probably fictional, but it's purported to be the capacity of a person to be able to recognize just by looking at them, whether or not a person is gay.

Let's suppose that that exists. It's a type of perceptual expertise. Is it a virtue to have that kind of perceptual expertise? Similarly you know somebody who can recognize just by looking at it. What a person's race is, what their background is etc. You know there's as Daniel Bernstein says there's no guarantee that perceptual expertise will have a net positive contribution to the proportion of true beliefs or knowledge.

And you know that assumes some kind of ethical theory where having more true beliefs or knowledge is a good thing that might be a consequentialist view. Might be. Social contracted doesn't really matter. We would have to spell that out, right? So many asks are privileged epistemic agents subject to different epistemic obligations than marginalized or oppressed.

Epistemic agents are, how might you ask that question in a practical sense? Maybe it should be incumbent on the people who are really able to distinguish between forms of discrimination say because they've been really well educated. They grew up without discrimination and are able to observe it in the world and all its forms because they get to travel.

Maybe they have more obligations than people, who just haven't been able to experience in the world in that way. Um, it's a good question. Course, that question could be recast as desperate, the smart people have more ethical obligations than less smart people, or do rich people have more epistemic obligations than marginalized or repressed, people, you know, our capacity may play directly into our ethical standing, and that's who a certain degree is part of the thinking behind say something like an ethics of a theory of care.

Where for example, a privileged epistemic agent, has an obligation to take into account and the marginalized person's inability to see their own repression and to say maybe help them see it or see it for them or something like that. To me. Interesting question. It all comes back to one of the models that these epistemic agents will have, are these models sensitive to the sorts of things that marginalized or oppressed people are not sensitive to.

And what role does that play and how they develop their own intelligence and how they conduct themselves. And if we build AIs that are similarly, much better able to distinguish between say good wine and bad wine, does the creator of that AI have an obligation to make that knowledge available to people who do not have that capacity.

So that they're not tricked into buying bad wine that the AI can tell is bad wine but they can't. Good question. You're making models. It's an art. It's a science. It's a practice. It's a series of choices and it's not just about the ones and zeros in the model, that's what we talked about in the previous section.

But now, in this section in this module, or in this video, we're talking about things like well, say, you know what problems are of high priorities. What do we want? Our AI detectors to be detecting? For how will the outcomes be used? How will we respond to adverse outcomes?

We touched on that during the discussion. Suppose we have an AI that is able to make ethical distinctions or suppose we have a society that is able to make ethical distinctions and it comes out with a verdict that you disagree with. It says something is ethical and you think not it's not ethical or vice versa.

How do we respond? What's the appropriate response? How even do we measure the outcomes of a model. We've trained our AI. It's got a model. The best evidence that it's the model. The good model is the training that we gave it, but we've already done the training. How do we evaluate for the appropriateness of this model?

That's a good question. Because all the evidence that supports the model has been used to train them model, it seems that there's nothing. That would prove that no model is wrong.

Models are trained, but the training is the results of extensive programming. We could. Look at that process, we could say, have they used rigorous programming standards. Have they done unit tests? Have they then usability tests is are all the HCI questions resolved and they then AB testing with actual users etc.

There's more is the program open source can be people see for themselves. How the application that generated? The model works. So these are all points raised in this article here that I quote, well, they're actually know, let's get that right there. All in sufort XL 2019 and this is a different article I think.

Yeah, this is what it's wrong with computation. Notebooks pain points needs and design opportunities. Choices. Choices. Think about, for example, how we would apply these models in counter factual situations, counterfactuals are special cases, and they drive people and semantics nuts. Because, you know, let's go back to current apps state space.

That's everything that could possibly happen in the world, so everything's covered, right? But a counterfeitual is something that hasn't happened yet and might never happen. The classic example is breakless trains are dangerous, which we can cash out as if a train has no breaks. Then that train is dangerous now.

It's a counter factual because there are in fact, no trains without breaks. And the reason for that is they're dangerous. Nobody would want to build one. But how do we know that a statement like breakless trains are dangerous is true? Where all the semantics says that we use a model.

And in that model we put a brakeless train and we see if it's dangerous, right? Kind of makes sense but that model can't be the real world because there are no brakeless trains in the real world. So it has to be, you guessed it a possible world. So we'll create a possible world which is exactly like our world.

But with one break. Let's train in it and we'll ask ourselves. Is that train dangerous? Well, first of all, do have enough trains in that world and where should that train go? I mean, should it be a train operating in Western Canada? Or should it be the London to Glasgow train or maybe the trans liberian railway?

Let me just lots of places. We could put that train for the. Principal is pick the possible world that is most similar to the existing world. Well, Now we got all of these choices that are outlined in this grid. What ontological perspective, do we choose to adopt and why?

So, it's a world, the world has trained etc. How do we know about the social categories? You know, there are train, drivers or train passengers. Etc. How do we know about them? What about our semantic choice? Are we going to use this possible world analysis that I just gave you?

Or maybe that's stupid and maybe what we should do is just say look we know what causes trains to do things that's create a separate causal model. It's not really a possible world model, it's just a puzzle model. That tests the hypothesis. How do you change between one and the other, what happens to the truth value of the counterfactuals of interest, you know, in other words okay you're saying breakfast, trains are dangerous.

I've created a possible world with a breakless training in it and I say that that's dangerous and I respond. Well how do you know have you been to that possible world you have access to. It could questions all similarity choice, what counts as the most similar possible world to our world.

But you read somebody like Robert Stoneacre. That's going to depend a lot on the salience of different features in that world but that of course, is going to depend on how we identify, what all the different features are in the world and if our worlds are ones in zeros, it's really hard to say that one set of ones and zeros is more similar than another.

And we could come up with a mathematical calculation of similarity. And that's what they do in a number of these algorithms that I talked about. Remember when I talked about the distance between two points on a graph but that distance is a similarity measure but there are different ways of measuring.

The similarities could be distance. It could be total. Number of features depends on. Hey, how you interpret it Context, how do these categories operate in the world idealization? You know we have a breakfast train in this possible world. We want to test it, we have to make it a random break list train about one that we designed because then we would just design the danger aspect right into the train.

So has to be a random breakless train, but what do we miss? If we do that and suppose our model actually let us say that. Yeah, maybe we should build some breakless trains in this world. Presumably, that would cause harms harms to the people. Riding the train. We might not say well yeah, whatever, but the people riding the train tend to be more poor when we look at our possible world.

None of our friends dieds. We said, okay, that's fine. And we didn't worry about the poor people, but maybe we should have maybe that should change our calculation as to whether breakfast trains or dangerous. See how much I can make of just one simple example, like breakless trains. Now imagine examples that are large complex involved, entire populations, you know, each one of these questions that we ask can be answered any of a number of different ways.

And any of these answers can have a bearing on the ethical output of our possible world. Some possible worlds might simply be unethical, you know? For example, we should not create possible worlds, that do not include entire segments of the population, it should be. For example, unethical to create a possible world that contains no people from Swaziland in it because, well, that's good argument, why would that be unethical?

Because people from Swaziland are important because not including people from Swaziland would would change the predictive reliability of our model, you know? Again, there's arguments upon arguments here, right? Of supposedly created our model. How do we validate it? What I mean, here is what problems will we, use it on where we'll use it, how will we use it?

You know what about the interventions that are provided based on the model? How do we validate those? Are we for example, suppose our model says that we should use a certain textbook for a certain student, there should be a mechanism to test. Whether that really was a good idea there should be standards perhaps for transparency.

What are the key individual independent variables? That our model looks at in order to make its determinations? And of course the outcomes assessment we've used this model for a year. What happened Since sort of things come up in learning analytics models? I've been talking pretty generally all along but in the suffered paper that I referenced earlier on, here's the full reference to it.

They provide basically a framework for learning analytics models. And you look at some of the look at some of the dimensions, they call them. And these are all dimensions related to, I would say, the interpretation of the model that would be used by the learning analytics. So they're using LA model here.

Differently that I'm using LA model when they're using LA model here. What they mean is the type of model is of, the perceptron model. Is it a boltsman? Machine model a GNN or a recurrent or network, right? But really remaining they. The access specific weights that was created by that type of model.

Either way, these are relevant. The pedagogical theory which will tell you, what kind of things you're looking for. Pretty good. Good example. Here is how our analytics system will measure the input that a person receives in an educational application. If our theory says, is based on a principle of say cognitive load.

Then part of what we're trying to detect is extraneous. Cognitive load, what counts as extraneous. Cognitive load will be defined by our theory. And my theory might say that anything not directly related to the learning outcomes is extraneous. Now, there are other theories that would not consider that extraneous at all because the output of learning is in the simply defined by the learning outcomes for a particular learning event.

But if we're using this theory, then we're measuring for extraneous cognitive load, and measuring the effect of that. The objective is the objective to meet the objective. The learning objected is the objective to achieve a high test score, is the objective to not drop out of a course. These again, impact, how we're going to interpret the model that we're using and so on right through data instruments, the competencies required, or that need to be developed by the people using the learning analytics models.

The constraints on the application of these models that include things like privacy, ethics norms, etc. All of these play into not just how we're going to set up the AI and the analytics, but how we're going to interpret the outcome, how we're going to interpret the model that is developed of each of the students taking part in the move, or in the online course, there is a domain of thought or domain of writing domain in the literature called model risk management.

Here, I'm referring to a seminal document from the board of governors and the federal reserve system. Describing this it's the SR 1107A1 that rolls nicely off the tongue, doesn't it and summarize very long document model risks may occur because the model may have errors or the model may be used and correctly or inappropriately and a risk.

A model risk management practice combines standards for the model development and implementation, some kind of model, validation processes and some kind of model, governance process, delineating, roles, and responsibilities for various stakeholders in the environment. One last thing it will wrap up this discussion. Just some suggestions from seabock which is a company of there in the field on modeling practices.

The four practices one stay in the operational field. This is important because models are trained within a very specific domain of discourse. You wouldn't use your dice model to predict what cards are you're going to find in a deck, they're two different sets of entities. Similarly, the models can be applied only in certain circumstances and even then only on subsets of that circumstances.

As they say here models can never simulate all the behaviors, a reactions of the system they operate only in one limited field with the restricted number of variables. Now there is out there in the world. This idea of general. Artificial intelligence where you have one like where you have like a human one model that handles everything for you.

In theory, that's possible and practice it probably won't be any smarter than a human. Unless it's got a lot of memory had a lot of power and therefore a lot of speed we're not looking at that in our lifetime and also it would have to go through a lifetime of training and we'd be right back to the original problem that we started with.

So, there are limitations certainly today limitations on how these models can be used. There is a field of study called transference or transferability or a model used in one. Domain can be applied at least in a general sense to another domain. It's easy to argue that that could never work but then, you know, disciplines like mathematics, if you develop mathematics and one domain, like, say physics, it may be possible to apply mathematics in another domain like say, engineering, it may be possible to apply it in another domain like, say, social science.

But, maybe not so much in the domain, like art. So, transference is a thing, but it's not clear how transferable these models are, secondly, evolved models. We shouldn't take any of our models, as a finished product because the world that their modeling, or at least what we understand to be the world that they're modeling changes on.

And so the model should be continually retrained or fed new data are, you know, have its parameters modified with new experience, etc. Especially within the models restart limits. We need to be continually developing new tools so that they can function properly beyond those limits. They recommend using several types of models.

That makes a lot of sense to me. The questions going to come up, of course, how do you choose between, which model you're going to use and then we get back to. Well, what are the different ways of validating models to we have? What would count is a proof as an of a good model?

Is it just intuitive? Are there actual numerical constraints we can describe and impose, you know? And and especially when we're thinking about, you know, what is the ethical outcome of all? How do we determine what the ethical outcome of a model is. Again, it's probably not going to be something that we calculate with numbers, unless we're a really hard core.

Utilitarian, it might be something that we see just by intuition, we look at it. We see. Oh no that's not right. Finally keep context elements consistent. And again, the context element is the environment in which the model is operating, if you're using a model and a learning management system, and the model has been developed for your geography course, you probably shouldn't change it to a history course on the fly, the reason for that is the results are unpredictable.

So that's what I have to say about models and interpretations again. This is one of these subjects that we can do in our course on there are there have been entire books written on this. There have been entire books written on single slides. And this presentation, nonetheless, it's certainly gives us a lot of food for thought about the sorts of decisions that need to be made in the design and the application of systems that create AI models and the environments that they work in in the field of learning and development.

So that's it. For now I just panicked because YouTube is complaining and saying it's not receiving enough video but I have a backup so we'll see how that goes and it's the end of the video anyways. So I can't do much about it. So I'm Stephen Downs that was models.

And interpretations next time I'll do a little bit about applying and evaluating models not very much on that, but I just want to raise some of those issues and then we'll wrap up this section. I'll workflow and move to the final module of the course. So thanks a lot for being with me.

See you next time.

Force:yes