Gunter Maris is a Belgian living in The Netherlands. He directs advanced psychometrics research for ACTNext.
In this interview, Maris explains the theory behind The Wiring of Intelligence, a network model of learning. He also quotes Leo Tolstoy and references Sherlock Holmes.
Maris describes the actions of a “crazy electrician” who wires the connections in our brains. The network theory of intelligence developed by Maris (with colleagues Alexander Savi, Maarten Marsman, and Han van der Maas) is based on the idea of a house with interconnected electrical sockets in every room that represent synaptic knowledge. Language, speaking, and reading have tight connections in the brain while mathematics skills are grouped together in another room.
For example, in the mind of a statistician, math connections can be tight and dense. Maris’ primarily Dutch-speaking brain has many “wired connections” pertaining to English terminology around psychometrics but probably fewer that cover cooking or gardening.
Connections from disparate rooms build measurable broad intelligence. The model helps explain the positive manifold theory (also called the g factor or general intelligence factor) first developed by Charles Spearman in the early 1900s.
Near the end of the interview, Maris explains how the simple model of networked intelligence is being overlaid with ACT’s Holistic Framework.
The views and opinions expressed in this podcast are those of the authors only and do not necessarily reflect the official policy or position of ACT, Inc.
Did you know?
You can listen to our podcast playlist on YouTube or visit the Navigator podcast homepage to listen to any episode.
[Gunter Maris] If raising a child is like designing an adult, what’s the purpose of the baby? It means the purpose and state of the baby is not determined when he or she is born. It’s something that emerges that we call growing up and that’s what this model allows for. It has literally a placeholder for educational intervention and changes in quantity and goal.
[Adam Burke] What is intelligence and how does it work? This episode of the navigator podcast is devoted to “The Wiring of Intelligence,” a paper that Gunter Maris and several colleagues published last year. They propose a simple model of learning based on a network of connections in the mind and that’s the basis of Gunter’s work: examining the process of learning and when education works or does not work.
[GM] I’m Gunter Maris. I’m the senior director of advanced psychometrics with ACTNext which is a research division in ACT. I’m based in the Netherlands whilst ACT is based in Iowa City.
To start off let’s go back little over a century that’s when the first intelligence tests were developed and around the same time in epidemiology, they started finding out that cancer is a disease that runs in the family. There are families with a lot of people with cancer and the question was why does that happen and epidemiologists came up with two explanations. One is cancer is a genetic disease and people in one family are genetically more homogeneous than between families and the second one is cancer is actually a contagious disease and people in the same family they spend more time close to one another and so that’s how it works.
They came up with two explanations for this phenomenon and they quickly found out these two cannot be distinguished just by counting the occurrences of cancer. Now around that same time intelligence testing started but we did something different. We came up with just one explanation: intelligence is the fact that you may give more or fewer correct responses on an intelligence test and it’s basically genetic. It’s because you are more intelligent that you give more correct responses to an intelligence test and as part of me for a long time and the wiring of intelligence out of that as means to try and explain why some people give more correct responses from a contagious process and that’s important because that’s something which is more malleable you can intervene in it.
If it’s in your genes literally there’s very little we can do about it. You were born to be stupid then you’ll grow up to be stupid and education is futile which is sort of a sad outlook. Whereas if it is contagious you can do something about it so we set out to collect a list of phenomena that we know about intelligence both at a population level, cross-sectionally everything correlates positively with everything else, also in a longitudinal within person perspective.
So intelligence tends to stabilize over time so we can ever better predict the future for an individual or for an individual the differences between people increase over time so people that are close together when they’re three years old are likely to end up further apart from each other when they get older and we build out a minimal theoretical model to explain these findings which is simple enough so that we can get new predictions from it. Then when we have a new prediction you go out and collect the data to test the prediction either it comes true and I’m happy, I go back to the drawing board and a new prediction.
Or the prediction does not come true and I’m equally happy because I go back to the drawing board and I need to revise the theory but I always end up getting back to the drawing board with my equations on it.
[AB] The wiring of intelligence is a network model and that means that things develop through connections.
[GM] Yeah, I’ll explain. Whether the wiring comes from literally comes from wiring as electrical wiring in a house. You can think of a person as house with rooms and there’s the crazy electrician who has to put in the electrical wiring. So he puts circuits in the room and the circuits you can think of as an item response, either there’s electricity on the circuit and response is correct or there’s no electricity and the response is incorrect or unconnected. The electrician wires these sockets together and if they are wired together they either both have electricity or they both don’t have electricity so that’s the whole mechanism of the model and whether a set of connected sockets have electricity or not we considered just to be purely random and why are there rooms in the house because everything correlates with everything else but some things correlate more with each other than with other things.
That’s as simple as mathematical or mathematics questions correlate more with other mathematics questions than they do with language questions correlate positively.
In mathematics algebra questions correlate more with other algebra questions or geometry ones and the rooms reflect that hierarchical structure and there’s a very simple mechanism, the crazy electrician puts in a little bit more connections wires within rooms than between rooms. That mechanism is enough to explain the full correlational structure you find in an intelligence test but also at a test like the ACT test.
[AB] And this is kind of the positive manifold.
[GM] That’s what they very fancily called a positive manifold and everything correlates positively.
[AB] So separate rooms have some effects on different rooms but more so within that room but then algebras say there’s a higher connection.
[GM] And this this wiring process is different for every person. There’s a wiring for your brain basically just one from my brain they are not completely independent, so the total set of possible wires is the same but are crazy electrician wired you and me and we are a little bit separate and what that gives us is that if you go across the population. It looks like we all have weak interactions between all the sockets but at the individual level that’s not true the wires are there, or they are not and that means that over time if new circuits are added and get wired to the system. Your behavior will be very consistent over time see what they call famine offense when the fence for the auto-correlation is high at the individual level you’re not like a randomly different person tomorrow that you are today and it’s wiring mechanism explains why that’s the case is because your wiring didn’t completely change overnight. Maybe one disappears it disintegrates. Some new ones are added, slight changes but in small changes over time but they change big differences between persons and the contagion mechanism in there arises from what his model predicts. Its members develop the predictors is that if two sockets are already indirectly connected so our third one or for a whole bunch of other ones. They are a little bit more likely to also get connected directly and what it gives you is that there is sort of like a big cluster of sockets that are all connected to one another and it’s a very stable structure because things that are indirectly connected get a little bit more likely directly connected.
That means if one wire disintegrates it’s still connected so that’s a robust structure and that’s the contagion mechanism and that backs out.
[AB] What I know about the brain is that if you remove part of the brain wires form around it and kind of make up for that absence.
[GM] Yeah, their interests access connections to make your brain it’s not directly modeled brain but to make your cognitive structure again some connections disintegrating over time otherwise you have a fragile brain if one of your neurons would disappear that you would no longer function that would be sort of an issue and this network forms naturally. So it’s constantly updating and making new connections and reinforcing connections but it does so in a malleable way that’s what we call education.
If you think of a vocabulary network, then I’m Dutch-speaking. I learned English but my network of English vocabulary is very dense a big in terms that have to do with statistics with psychometrics. I don’t know a whole lot of English words having to do with gardening or cooking so it’s specialized in a particular area whereas somebody else who is maybe also Dutch speaking but who happens to be a cook working in the US and at least our restaurant that person will have a very large vocabulary network with lots of intricate connections having to do with cooking but he probably wouldn’t know what a point by sale correlation is and that not only goes for vocabulary but for all things so you can decide or intervene in how to network those there’s no preset and state you literally built the end state.
In a meeting a while ago I summarized as follows: if raising a child is like designing an adult what’s the purpose of the baby? Of course, everybody laughed about that but there is a serious part of it.
It means the purpose, the end state of the baby, is not determined.
When he or she is born it’s something that emerges out of the process that we call growing up and that’s what his model allows for it has literally a placeholder for educational intervention and changes in what the end goal is whereas the genetic idea about intelligence is more like a child is more like a cannonball which we aim at conception and then we shoot off the cannonball that may be some randomness due to wind or something but where the cannonball will to end up is pretty much set in stone that’s also pretty sad perspective for education.
Whereas our perspective you have the potential to intervene to change the course to go back if you find out that some link which you would like to have been there is missing you can go back and mediate the missing link literally.
[AB] Have we explained the positive manifold and how that relates enough do you think?
[GM] Yeah, it’s a necessary consequence of the theory that also means that it cannot be negative so correlations have to be positive, being better at something has to go with being better at something else not with being worse there’s also a difference with the more traditional factor analysis models essentially of intelligence testing they allow for any correlation between anything but we don’t just find any correlation we always find positive ones and this theory says it has to be positive so if I have a mistaken idea that affects negatively is that true if I if I have a wrong conception of part of the world that will go with a particular group of connected things that depend on it say if you mistake addition or multiplication for addition you will give a structured set of incorrect responses.
You will say 9 times 9 is 18 right and 3 times 2 is 5. So that’s that gives positive correlation because these long responses go together but because that’s something at the individual level that cannot happen for everybody so 4 over the whole population the correlations are positive whereas within the individual that can be groups of things which are all negatives you to a misconception that can be fixed and then they would all flip. So if I talk to the child and say no you don’t have to add the numbers you have to multiply and work a bit a bit she sees the light then all of a sudden 9 times 9 becomes 81 2 times 3 becomes 6…
[AB] How did you start on this wiring of intelligence? What was your inspiration or who are you studying or take us through the process?
[GM] It started with the academic epidemiology problem, with the realization that the models we’ve been using for intelligence and for education in general, they are based on there is distinct ability that you have which in very strict situations is genetically encoded, you can’t change it and that causes your responses to be correct or incorrect and that makes education futile that doesn’t it doesn’t add or fill together it explains large-scale educational testing large field intelligence testing as cross-sectional data but it cannot explain individual learning. I mean it’s pretty clear that learning does take place and we started out almost 10 years ago thinking seriously about that and why 10 years ago because if you want to study learning you need to be there when it happens and having a researcher looking over the shoulder of as we get in the K-12 area all day long is like a pretty heavy lift but computers do help. So online learning systems, they do have the potential to have us virtually be there while the learning takes place and then you got into a chicken-and-egg problem. You need to have somebody what you’re going to do with those data before people are going to be willing to start collecting them so this effort is part of that to get to a theoretical understanding what are the things if we would or should be looking at or could be looking at once we start getting data on large numbers of learners with a very high frequency like daily.
[AB] What does the test look like or how do you gather this information?
[GM] Well, one of the predictions we are currently in the process of testing is that summarized by the opening lines of Anna Karenina by Leo Tolstoy:
“All happy families are happy in the same way. All unhappy families are unhappy in their own way.”
Now if you replace “happy” by something, like dyslectic or depressed or something else, that’s a prediction that comes out of this theoretical model that the cohesion the correlational structure between symptoms of this lecture should be lower for people who are dyslectic. They have more symptoms more spelling errors more reading mistakes but they all have their own individual pattern of this lecture and as a testable prediction, if it’s true, it corroborates our model if not we need to revise the model and it has direct implications for remediation because it means you need to deal with every patient of this lecture but also of depression individual and there’s no one-size-fits-all therapy for dealing with dyslexic students like education, there’s no one size fits all.
It has to be personalized that’s why we’re in the process, what you need for that is at the minimum data from a sample of students some of which you know like they to otherwise what else you ignore. They are not dyslectic give them a test and look at what the correlations between these responses look like and they’re supposed to be lower for the dyslectic.
[AB] So how do you measure the network of intelligence?
[GM] The basics because these two classes of models are statistically equivalent or one is a subclass of the others so sectional measurement doesn’t really change everything that goes with the item response theory model still goes with this network model but over time it’s a different thing and at the individual level is different because if you think about the electrical analogy finding two sockets who are in a different state, one has electricity the other does not, tells us that there is no connection between them if they both have electricity we don’t know whether they are connected it could be could not be but if they’re in a different state we know for sure they are not connected and if we think there should be a connection between these two concepts or questions we can intervene on that. So not having learned a connection is something we can observe we cannot observe whether you have lowered the connection that’s the limit of what you’re doing. That’s about the theory as it currently is stands.
[AB] You don’t know what you don’t know kind of thing. One of my questions here is what are the limits of your network intelligence, the wiring of intelligence?
[GM] Well, as a theory intended to put an alternative on the table but if you think about how children take an actual intelligence test or an ACT test the response they give it comes out of a process and that’s not the way it’s in the current theoretical model there are four responses to test items there are no direct indirect interactions between the questions that’s different than other fields meaning psychopathology if you are depressed it’s highly cool possible that there are direct relations between the symptoms because they exist also if I don’t look at them but fortunately in your brain every question that has ever been on an ACT test does not exist at this moment. In your brain the response emerges out of a process that starts when I ask you the question and we know that we need to move in that direction. So this is certainly not an endpoint is more like a starting point and to get attention for the idea that there could be different explanations for what we see in large-scale educational or intelligence testing than these latent variables and so we need to work in that direction but for that we first will need the type of time series a data at high frequency to start studying because it’s like in Sherlock Holmes: It’s a capital mistake to start theorizing without data.
So we first need facts and then we can find out more about what the actual response process looks like out of which the response you give emerges and how these processes relate to one another for different questions across the last four times six something happens in your brain that gives the response and if I ask the question two times three something happens as well and it’s not orthogonal to the other one the process is already latent and there is at that level probably more sort of a wiring mechanism but at this moment we mainly have a lack of high quality data so that we can find out what are the regular like regularities any in all these individual learning processes process the dose we can add to the collection of things that we know and I mean need to start taking about what is the simplest non-trivial model that implies all of these phenomena that’s the way you build a theory you take facts. We always see these twenty facts what is the simple story that explains why we see these twenty facts and nothing different they have to happen and now we first need to wait until we have 20 additional facts that relate to actual learning and change and the effects of intervention and other things and with those new facts we go back to our drawing board say. Given those facts what is the simplest non-trivial theory that explains why is taking place and this is where you’re talking about the you want to observe day to day even maybe minute by minute yeah the process of making connections yeah you want to be there at the moment when a child realizes something and you can you know when it happens what the kid was doing. If you think of a big learning system with educational resources and quizzes you see if kid watched this particular YouTube video explaining this particular concept then with high probability it clicks and if you quiz master to see they caught on to it and if they catch on to this then they can move ahead to a more complicated concept.
It’s like in biology, if you have to understand osmosis. You need to understand diffusion and you need to know what a semi-permeable membrane is. If either of those two are missing, you’re never going to catch on to osmosis as a process and those are things we want to be able to observe.
This kid probably did not catch on to sufficient level what diffusion is and so there’s no point in giving in another video or lecture on osmosis. We need to take a step back fix this problem and then we can see whether that help in getting an understanding of osmosis and as we have a collection of possible interventions say read the chapter in a book watch one of five videos talk to your teacher and we have a sufficient number of students you can figure out which one works best and that allows you to optimize education also at a personal level you can find out maybe for boys this video is bad and for girls this chapter works better.
You can do it at a fairly fine-grained level but it allows you to optimize the interventions by that you can experiment and get a grip on how the process actually works and that’s also goes back to epidemiology. From just counting the occurrences of cancer you can’t distinguish the two mechanisms yeah but if you think of a disease like Ebola and suppose there are a bunch of outbreaks say ten of them four five of them you guarantee the people who have the disease as a five you don’t within two days you know whether Ebola is contagious or genetic. That’s the way they do it we need to find a similar way to do experimental intervention because that does allow us to tease part these two possible explanations which with the data we have today.
They have the same explanatory power but if you can intervene in the system you can find out whether it is contagious but as genetic or likely some makes others too.
[AB] We talked about the limits of your model and you use this George Box quote: “All models are wrong but some models are useful” and so how is your model wrong and how is it useful?
[GM] I explained a bit on how it’s wrong but it doesn’t take into account process that there are likely not direct interactions between responses to test questions but model is useful and that it’s highly simplified in the whole model there are two parameters one relating to the probability of getting a wire we’re gonna do and another one a little bit lower between rooms so it doesn’t get much simpler than that but that allows us to derive detailed qualitative relations from it like this Tolstoy quote we talked about before which you can test with data. We have a very detailed model with lots of parameters. It’s very difficult to get new predictions out of it because the more degrees of freedom you add more parameters the more things you are allowing to occur in reality and at the end of the day you don’t explain anything anymore. You may describe very well what’s happening but it’s not very easy to make sense of why does this happen and it’s also highly likely if you keep adding parameters that there are multiple ways in which the same thing can happen. That’s why the model is wrong but we don’t care about it but it’s useful because we can get new predictions out of it which we can test to Levis the model and replace it with an probably equally simple but slightly different model.
[AB] What’s next for your project will you be using computer-based testing?
[GM] Yes, next is are a whole lot of things one is testing predictions on this particular version of version of the theoretical model the other one is feeding into the theory information that we have about what the structure of cognition is a still about the HDF hierarchical framework and that’s basically a categorization of what the rooms look like.
[AB] The holistic framework?
[GM] Yeah, so you have an office there are larger areas that’s big rooms in it are cubicles which are smaller rooms. What the holistic framework maps out is which things in mathematics are most closely related to which other things and do that at various levels. So you have linear equations and quadratic equations they belong algebra, they belong to mathematics, they belong to the curriculum in the end and it looks much more fine-grained and that information we can use because that tells me what the potential collection of wires is and everybody gets sort of a subset of these but at least we now have an idea what the potential structure looks like and then we can build a more concrete and association of the model that’s being developed currently. It’s called the ACT master model which is closely related but it’s not just an abstract model. This more concrete one that you can connect to real data learning and assessment systems and my group works on theories statistical models and methods has statistical algorithms to support learning measurement and navigation that gives like a nice 3×3 table with nine cells in it and a problem with what we know currently there is no one coherent approach that connects all these nine cells together. So the overarching theme is to come up with theories models and methods that work for learning measurement and navigation that’s the work in progress will take a while to finish that. At this moment we are mostly working on the learning and measurement part and that’s mainly because we have more information on that more empirical data and an application part as intervention in the writing system for that we need to get a high quality up and running learning system at scale so we can also bring that on board we want to just fill these nine cells so that it all fits together as one system.
[AB] Good, we can stop there and thanks Gunter.
[GM] You’re welcome, it was fun!
[AB] That’s our show. Thank you for listening and thanks again to Gunter Maris for joining us.
Have you heard about edu2050 and the creative thinking test being developed for PISA 2021? Find out about ACTNext prototypes at ACTNext.org/prototypes and learn about our latest projects and ongoing research at ACTNext.org/projects