Downes Presentation

From LTCWiki

The Recognition Factor

Stephen Downes

Online Connectivism Conference Day 5 - February 8, 2007

Contents 1 Abstract: 2 Introductory remarks 3 Practical Examples 4 Distributed Knowledge 5 Patterns and clustering 6 Pattern Recognition 7 Questions and responses

Abstract:

Connective knowledge is based on pattern recognition of emergent phenomena in networks. In order for a pattern to have any meaning, therefore, it must be recognized. This means that knowledge formation in a connective environment is a combination of two elements: the perception, which is the pattern to be recognized, and the perceiver, who does the recognizing. Knowledge, therefore, is not uniquely inherent in a network, but exists only insofar as it is recognized to exist. This talk will explore this argument and its implications on a theory of connective knowledge.

Introductory remarks

Hi, I’m Stephen Downes. I’m online live right now, in the OCC2007 conference. I’m just listening to George as he speaks; he’s just introducing me now. In a few moments, I’ll be beginning my presentation to the OCC conference. There may be a few minutes of silence here as I wait for George to finish and then you’ll hear the session get going.

Thank you, George. Thank you. Hello, everybody. A pretty good-sized crowd here, which is appreciated. Thank you for showing up on Thursday – for some of you morning, some of you afternoon. For those of you in the Eastern time zone, Thursday at lunchtime, which I’m very impressed by. The title of my talk is “The Recognition Factor.”

• Practical Applications
• Distributed Knowledge
• Patterns and Clustering
• Pattern Recognition
• Reliable Networks

The talk breaks down into five major sections, and George, what I will do is talk about each of these sections for ten minutes or so, give or take, I’m going to try to aim for that, and then we’ll have some discussion after each one of these sections, if that sounds all right. I will also from time to time as I’m proceeding through the sections, turn off the microphone and pause for a few seconds in case anybody wants to jump in at that point.

That was an example of me pausing, and especially, George, you should feel free to jump in at those pauses, and other people if the spirit really moves you, keeping in mind that we now have 111 participants and a limited amount of time.

Practical Examples

Where I’m going to begin is the context, at least for me, the context of some of the discussions that we’ve been having online recently about whether connectionism is a new theory.

What I’m going to say is that a lot of the stuff that I have to say in my talk is stuff that you already know, things that you’ve already seen. You listen to some of the things that I would have to say about learning, you’ll say well so-and-so said that, so-and-so said that, and that is quite true. What I’m trying to do with this talk in particular rather even than construct some sort of argument or inference, I’m trying to get you to see it differently, to see it more clearly, so that when you go back and you look at the everyday things that I talk about, or that George talks about, or that other people talk about, you’ll see them differently, you’ll see them in a different light.

For instance, you should see in front of you, assuming that Elluminate is working, a nice little green man standing on a planet looking at the sky. This is a scene we’ve all seen, right? Night sky, sun’s just sort of setting off to the right there, stars are filling the sky. And we ask which ones are the planets? If I talk about stars, if I talk about planets, you say well, you’re not talking about anything new.

But if you change your perspective, you look at things differently, from different light, you tilt your head, then you can see the patterns underlying the picture. We’re standing on the surface of a planet, a round planet, at about 55 degrees, that’s the angle of the tilt there, and so the horizon, we can sort of picture ourselves on the edge of this planet, and we can imagine in our mind the orbits of the planets around the sun. You see the sun, now it’s setting off to the right. Now the planets are now easy to spot, because all the planets orbit the sun in a plane, the orbital plane. So you just pick out the bright little orbital dots on the plane. Now the idea isn’t to teach you how to spot planets, although you’ll find that a useful skill in the future, I hope. The idea here is just to show you that even something very familiar that’s right in front of you can look different if you look at it in a different way. This example, to give proper credit, comes from Paul Churchland.

My theory of learning is very simple. My theory of learning is: to teach is to model and to demonstrate and to learn is to practice and to reflect. There’s no short cuts here, no magic solutions, no magic wands. You’re not going to get more of it if you throw more funding at it, less of it if you throw less at it. It’s just a very simple theory, and that’s, in my mind, because brains are at heart very simple. And of course it’s something you’ve seen before. But to get to the heart of it, we ask: To model what? To practice what? What is it a person is supposed to be doing? And that’s the topic I’m trying to get at, that’s what I’m trying to draw out and to get you to see.

Here’s a little example. Suppose you want to teach people the concept of brackets in mathematics for some reason, and we have an example. There’s the example up there on the blackboard. And I ask you: to teach the concept of brackets, would you use this one example over and over? And pretty clearly, no, you wouldn’t use the same example over and over. Nobody would do that; that would be insane. But the question is why not? Because if teaching is nothing more than the transfer of knowledge from one person to the other, we should be able to manage it just by doing the same thing over and over. But of course what we’re trying to do here is to teach a concept, not a fact, and a concept is something that is deeper than what you see in any given example. It is something that underlies the example.

The question here now is the concept for bracket or spelling or for architecture or whatever, best thought of as a rule that we can teach people – we just tell them what the rule is and they follow it – or is it a pattern? And, as I’m sure you can imagine from the title of this talk, I’m going to argue that it’s a type of pattern.

I’m going to pause here and I know I don’t seem to have covered a lot, but I still want to pause here and see if there are any comments.

Distributed Knowledge

And I see we’re getting some great comments here and questions in the chat area, like can a concept or pattern be thought of as a rule.

And what I’m going to argue here, in this section, section number two, is that there is a very fundamental difference between things that are like patterns and things that are like rules. We get at that by the concept of representation. What we’re most used to as far as representation is concerned is the model I’ve got here on the picture in front of you. I have the word “tree,” and that stands for a tree. And what you’re supposed to notice here, first of all it’s symbolic; it’s a symbol system. Secondly, we have a one-to-one representation here. The word stands for the thing; the thing is represented by the word. Now you can have one word standing for a class or a group of objects, but the point here is that the word here is atomic. There isn’t something that is part of the word “tree” that continues to represent something. You have the letters, but letters don’t stand for anything by themselves.

Distributed representation, on the other hand, is something that is very different. You have the same tree. We look at the network there that we see in the upper right, and the tree is represented not by something atomic, something that is a symbol, but rather by a set of connections. I’ve highlighted for the example here the connections in red. What you want to notice here is that there is no one thing that is the representation of “tree.” There’s actually one, two, three, four, five, six connections and looks like seven individual units or neurons there.

Also, and this is really important, the very same network with the very same set of connections, and the very same units or neurons actually represent various things. So we have the representation of the tree, as we had before, but the very same collection of neurons is a representation of a puppy – a nice, cute puppy – that’s what turns up on Google when you search for puppy – or a couch. So if you think of a sentence-based or symbol-based representation, the word for tree doesn’t mingle in with the word for couch at all; they’re two separate things, so that the meaning of the word tree is distinct from the meaning of the word couch. But on something like this, if you change the meaning of the word tree, it’s also changing the meaning of the word couch, even though semantically they are not related at all. So there are in a distributed representation, connections or associations between concepts that would also seem to us to have nothing to do with each other.

Here’s the theory: concepts are not words and that’s why it’s not going to be a rule-based system; they are patterns in a network and that like the human brain or a network like society as a whole. In these networks, there’s no specific place where the concept is located. The concept is distributed as a set of connections across the same network and other concepts are embedded in the same network; they form parts of each other and they affect each other.

So that’s part two, and I’ll pause here to see if there any comments or questions. I am going through this a bit faster than expected, but I think that’s OK.

Questions on the comments that are coming up.

What I’m going to do is defer that question until later on in the presentation. Right now I simply want to establish that concepts are a pattern, OK? The short answer – I guess I’m not deferring it, am I? – the short answer is that the principles that govern the operation of rules are sentential principles, they are logical principles, they are principles of symbol systems. And these principles are different than the principles that govern operations of patterns, and to express that sort of difference in a slogan type form, symbol systems are based on the concept of equality, identity, this is that. Patterns are based on the concept of similarity, this is like that. And that’s the short answer.

Right, and as was indicated in the comments – not the comments, but in the discussion in the Moodle threads – constructivism, to my understanding of it, is based in physical symbol systems. You don’t get constructivism without symbol systems. And since this is not symbol system theory, it is not constructivism.

My take on this, and we may have to bat this around a bit, George, connectionism is a computational theory that has been around for quite a while. It’s the principle of how patterns in networks are recognized, and then we can go on and start talking about that. Connectivism adopts all of that; all of connectionism is absorbed by connectivism. And then connectivism becomes the study of how we use this knowledge, this knowledge of connectionism in a pedagogical sense. I’ve represented it in the past, although I don’t think you liked this: connectivism is the pedagogy of connectionism. That would be my take on it right now, but at heart I’m a connectionist, my roots are in connectionism.

That’s not a bad characterization. I’m not sure that that’s the distinction that would have been intended, but that’s kind of how it happened. A lot of discussions about networks that occur in society fall under the heading of discussions of social networks, so you get people like Duncan Watts and others and you’re looking at stuff like scale-free networks and all of that. And you’re quite right, connectionist literature focuses on what they call neural networks or simulations of neural networks and it’s focused on things like the brain and perception and recognition by computers and so on. Part of my position is that the two phenomena are one and the same, that what we’re seeing at the micro level in the brain is the same kind of thing that we’re seeing in society, that we’re seeing in different ways in different places in society. The same principles that govern crickets interacting with each other govern bloggers citing and quoting each other, govern the development of river systems and trees – those principles are also the principles that govern things like human brains and computer networks set up in certain ways.

Now, the interesting thing, and the reason I’ve taken the tack that I’ve taken in this particular paper or presentation is that we have a network here and a network there, a network there, we have a brain, we have a society, and how do concepts go from one of these networks to the other? This is where we get to connectivism. The whole point of teaching is to get some sort of knowledge, if you will - whatever you want to call it - from the big network, that’s society, to the little network, that’s the student’s brain. But at the same time we’re also working on the problem of how to get stuff that’s in the little network, say my brain, into the big network, society. And so we’ve got this process that goes back and forth here, and what’s really important – it turns out to be crucial – is that this is not strictly a causal process. You don’t just simply cause a pattern to occur in a straightforward X causes Y sort of thing. Me having a thought in my brain and then saying it publicly like here, doesn’t automatically cause that thought to be part of the wider social network.

So we need to understand what that process is, how concepts go from network to network, and the first or beginning steps of that is to understand what a concept in a network is to begin with, and that’s why I’ve taken the approach I’ve started with here. So we get to the point concepts are not words. You’ve got to get away from that idea, talking about concepts as though they are words; concepts are patterns. So the properties that concepts are going to have are not going to be logical or syntactical properties; they are going to be network properties. So that is how I would answer that, George. How did that work out for you?

That’s exactly right, George. I’m going to try to drop my volume a bit. I’m getting into the red zone. There, how’s that? Yeah, I think that’s better.

Patterns and clustering

Now the way our discussion went, we started talking about the properties of networks. So what are those properties? The significant property that I want to talk about here is self-organizational or organization or structure. I’m tossing various words around, trying to get you to have a picture of it in your mind. I’m quoting here from Christian Hubert: “Self-organizing systems acquire new structure without specific interference from the outside. They exhibit qualitative macroscopic changes such as bifurcations or phase transitions.” It’s kind of interesting if you have any interest in astronomy, at the moment of the big bang there was one of these phase transition things. So at the very edge of the galaxy you got this ripple pattern, which is the ripple that occurred in the big gang. And why did it occur that way? Well, it just did.

The ways things connect with each other is reflective of the properties of those things. Different things connect in different ways. Here we have a diagram of various molecules composed of oxygen and nitrogen and S – what is S, sulfur? Silicon? I’ve forgotten. So oxygen will connect to nitrogen in certain ways, but both connect in the same way to hydrogen, so we get distinct kinds of networks. People talk out there about networks as though it’s a unary phenomenon - the network has such and such properties - but different networks that are composed of different things have different properties.

Networks don’t obey syntactical rules, they don’t obey things like odus ponens, which is the Latin name for “if a then b,” “a, therefore b.” They don’t obey principles of set theory or anything like that; they obey the laws of physics, and the laws of physics present themselves in interesting ways. For example, these forced patterns in this construction.

The organization of network is influenced by external stimuli, and that’s why we get the whole concept of perception and learning and all that. The actual structure of a network, the actual way a network is organized, the physical connections even between neuron and neuron in the brain is impacted by external phenomena. Now that’s not unique to human brains, if you think about it. A river system is a network. One of the external phenomena that affects the structure of a river system is rainfall. If you get lots of rain, you get more rivers; if you get less rain, you get fewer rivers. Another external phenomenon that impacts river networks is earthquakes or dams built by humans – all of these sorts of things. So the structure of networks can change according to the influence or impact of external phenomena on networks. When human brains change their structure because of external stimuli, this phenomenon is known as plasticity.

So we get many different types of networks. The network that people have been talking a lot about in relation to the Internet is something called a scale-free network. A scale-free network is a network that forms a pattern, types of patterns pictured here. You see a river system here on the one hand, and then you see the other scale-free network in the lower left-hand corner. It’s the type of network that has these spokes and hubs and you have one large hub and many spokes coming out of it. The Amazon River there, you have the Amazon River, which is really big, and then you have all of those tributaries that come off the Amazon, and then you get to smaller and smaller tributaries. That’s a typical kind of scale-free network. On the Internet, people like Clay Shirky and others have been talking about networks, hub-less type, and describing them in terms of power laws. A power law is a law, a diagram that describes the way networks look, and it describes networks where you have, one say, person with many connections, like an A-list blogger like, say, Clay Shirky, and you have many people very few connections, like, say, the rest of us. And all of this discussion about power laws has been depicted as “that’s just the way networks are,” but what I’m trying to show here is that is not just the way networks are. There are different networks that are the result of different phenomena: the properties of the entities, the external influences, physical laws.

Different kinds of networks, moreover, detect or emulate or produce – again I need to use these works kind of vaguely – different kinds of patterns. If you have a single-layer network, just one neuron and another neuron, you can detect binary phenomena, a or b, something’s off or on. Two-layer networks can detect open or closed regions. Three-layer networks, as the diagram shows, arbitrary forms or shapes. Again, you have different kinds of networks, and now you have different kinds of networks detecting different kinds of patterns. What that means is if what we’re up to is pattern recognition, we build the network one way, we’re going to get one kind of pattern; we build the network another way, we’ll get a different kind of pattern.

So I’ll pause here just for now.

That’s a great question. How do I want to phrase this? She’s asking are networks sentient, and how do I want to phrase that? Let me bracket that and come back to it. I want to say, Karen, that they can be, but by sentience we mean a very specific sort of thing, which means that some types of networks, networks with certain types of properties, will exhibit the quality that we call sentience, but not just any network will exhibit that property. A bunch of crickets chirping together, probably not sentient.

Let’s defer that. The question – because one of the questions that came up was do networks detect patterns by themselves? That was again part of Karen’s question, right? And if yes, are they sentient? Two separate questions, right? So I’m going to want to say that yes, they do detect patterns by themselves, but there’s a big but here, and the but just is the name of this title, “The Recognition Factor,” so maybe I’ll move some of those questions into this next section and talk about recognition and the role that recognition plays.

Synchronous means it’s a kind of property that occurs in certain kinds of networks. If you look at a network, analyze a network, you have one unit, you have another unit, and they’re connected. One unit fires off. The question is does the other one fire off? You get synchronicity if there is a propensity for the other unit – sorry – if the propensity for the other unit to fire off is greater given the first unit firing off than it not firing off. Take crickets chirping. If a cricket is more likely to chirp when it hears another cricket chirp, then you’re going to get this kind of synchronicity. If crickets just chirp whenever, you’re not going to get this synchronicity. Now for those who have studied networks, that’s a bit of a gloss, but it’s kind of like that. So the question of whether you get synchronicity in networks fundamentally comes down to the properties of crickets, what kind of thing crickets are, what kinds of entities in a network crickets are. That’s essentially the answer to that.

What I think is interesting though, the synchronicity in a network, all the crickets chirping, all in the same pattern, is nothing more or less of anything, it’s nothing in itself. And that’s what I want to get into at this particular junction. Let’s look at the network that forms the human brain. Again, the nature of the human neural network means that we create patterns in the mind in certain specific ways. You look at the neural network for visual pattern recognition. We’re set up in such a way that when we get visual phenomena, we find patterns in that visual phenomena – that’s just the way human brain neurons work.

And in particular, the way they work is edge detection. We see objects, we distinguish objects one from another because that’s how the layers of human cells connect to each other. Think of it as a representation. The sixth or seventh layer of the retina, what that layer is doing is finding a two-and-a-half-dimensional representation of a physical object. The thing is, though, in the case of pattern recognition, there’s more going on than simply the input phenomena and then the process of that phenomena to produce patterns.

You look at the examples here, the face of Jesus on Mars, or the face in the - I forget the name of that cloud that was detected by Hubbell, even the Virgin Mary on the side of museum there. We see patterns - our mind our pattern recognizers – but we see patterns in everything, whether or not the pattern is actually, shall we say, in the thing. What is happening here, and George, you alluded to this earlier, we, the recognizer of the pattern, are bringing something to the equation. Pattern recognition is not simply the matter of input phenomena at a given point in time.

Here’s a very famous example. Many of you have probably seen this. What is this? Is it a duck? Or is it a rabbit? It is a perceptual phenomenon that has been placed in front of us, and so we’re seeing a pattern and our minds are divided – which way do we go on that?

If we interpret that particular phenomenon one way, one set of connections becomes prevalent in our mind, becomes more active or more activated, it’s a duck. If a different set of connections becomes more activated, it’s a rabbit. Our interpretation of whether it’s a duck or it’s a rabbit depends on the prior organization of our neural net, depends on the prior organization of the network.

The network has a tendency to interpret things one way or another. Remember back from the earlier parts of this discussion, all of the concepts that we have are all sitting there in the network, they’re all overlapping, interweaving with each other. So when we’re presented with a new phenomenon, we have a tendency to interpret it in one way or another way. We can describe this as energy states, various neural net configurations. It’s like our perception of something is attracted by one of our prior or previously existing conceptions.

Associative memory is equivalent to patterns of connectivity, which is the creation of attractors, the creation of prior candidate patterns, which is in a word, recognition. To put all of this in slogan form, and keeping in mind this isn’t slogan form, knowledge in this theory is like recognition, learning is like perception. The acquisition of new patterns of connectivity, new attractors, new ways of seeing new phenomena through previous experience.

Now this is a phenomenon you’ve seen before. This is nothing new. You see this every day, when our brain snaps into one of these patterns. And Tom Haskins is picking up on some of the discussion that we’ve been having in this conference. He puts all of this in a post that he calls emergent learning. And all of these things, the "Now I get it," “A-ha!” or "It came to me out of the blue," "My mind leaped," "I did an about-face" - even like the conversion of Paul on the road to Damascus - these are all pattern recognition phenomena.

Knowledge is recognition. It’s like a belief that you can’t not have. It’s like after you’ve found Waldo in that picture, you can’t look at that picture and not find Waldo again.

Knowledge is like recognition
Learning is like perception
the acquisition of new patterns of connectivity through experience

I’ll pause there.

Do we have a clear understanding? I wouldn’t say that there’s a generally accepted consensus. Here’s my take. There are different types of neural networks again. These are described in the connectionist literature. I’ll describe three types of networks. One is the simple Hebbian associationist network. That’s where we have two neurons fire at the same time, the connection between them is strengthened. If they don’t fire at the same time, it’s strengthened. But if one fires and the other doesn’t fire, then the connection between them is weakened. So that’s your pure, inductive perceptual knowledge; you’re just drawing associations from the phenomenon.

Another type of neural network learns not simply through Hebbian association, but through what is known as backpropagation. Now this kind of neural network presumes a desire to respect an end state. So what happens is, you have a perception, you get input into it, you form your connections, you create your output, and then your output is either accepted or rejected. If it’s rejected, then this rejection propagates back through the network and weakens all the connections that led to it. A simplification of this might be the pain response. You see, “ooh, fire,” your brain goes ding ding ding ding, forms a set of connections, and the output is you that touch the fire, you feel pain. Now you’re getting back propagation and all the things that led you to touch the fire, all those connections, are weakened in your mind. That’s a second model.

A third model is one known as a Boltzmann neural network. And the idea here is that you have all your connections, your input comes in, your connections all form, and it’s like you drop a rock into the pond and the ripples flow back and forth, back and forth, and then they settle in, the pond gets flat again. The mind does that sort of thing as well in this theory. You get a perception, your mind gets all agitated, and so now you got some neurons are strong, some are firing off like crazy, some connections are strong, your mind is all out of balance. You go through a process where your mind settles into a stable configuration, except that it can’t just simply settle into a stable configuration, because the stable configuration that it settles into might be what is called a local minima – it’s a stable configuration, but it’s not the most stable configuration. So you agitate the mind, and the principle would be you’re sitting there, resting, you’re reflecting, what you’re doing is shaking up your mind, like a dream experience even, if you will. You shake the mind and it settles again into a more stable configuration. That’s a process known as annealing.

My expectation, and I think this is reasonably supported in the literature is that the processes that result in the organization of the human neural net is a combination of all of these processes. That’s why when I say to learn is practice and reflection, it is practice and reflection. You need not just the associationism, not just the backpropagationism, but you also need that time where you’re settling into a stable state; you need that annealing time.

Pattern Recognition

OK, cool. So what we’ve got here is a theory where knowledge and learning is a process of pattern recognition, and no this is not slide 30 of 100, it’s slide 30 of 31. No, I’m just kidding. Pattern recognition isn’t one of these hard and fast things. Pattern recognition isn’t, as I said in that slogan form earlier in the talk, it isn’t based on identity, it isn’t based on equivalence or equality, it’s based on similarity: something is more similar to something than something else. So you recognize the pattern nine in the phenomenon; you’re comparing the phenomenon with other patterns you have in your mind. In this case, the previous pattern is zero through nine, and the current pattern is most similar to the one we represent as nine. And what we want when we’re talking about learning is for students to recognize patterns in existing networks; for example, in communities of practice, communities of experts, the sorts of things that we talked about as substantiating knowledge in society. We want them to be able to recognize patterns in those networks. So we want the student to look at a phenomenon in society – say a physics experiment, a discussion among physicists, indeed novel phenomena, things that they haven’t seen - and be able to reliably extract patterns from that.

We want to go from the community, and here I’m pointing – you can’t see me, but I’m pointing – at those three scientists there, those three stereotypical caricature scientists – community to pattern recognition to that “aha” moment, to where our understanding snaps into place. We see the pattern: we connect what we see in front of us with some knowledge or some pattern that we already have.

But the thing is, as I said, different networks produce different patterns. So we can put one kind of network in front of students and get one set of results, we’ll put another network in front of students, we’ll get another set of results. And for that matter, we’re in this network. We are the experts that are the network of experts, and we want to produce knowledge in our network, we want patterns to emerge in our network that are good patterns. We’d rather come up with the theory of gravity than, say, the theory of impulses or impetus. We want to be able to maximize knowledge.

So what kind of network do we want to build? I’ve drawn two extremes here, and this was the aha moment for me. I remember it was 1989 or 1990, and Francisco Carella was giving a talk on immunology at the University of Alberta Hospital and I attended that talk. He drew a little chart on the wall, connections among entities, and you get too little connection and information never propagates; but on the other hand, you get too much connection and information propagates too quickly. If everything is connected to everything, then we are overwhelmed, we assume anytime we get a new perception, it overwhelms everything else. What we’re after is a type of connection that will fall somewhere in the middle, a type of connection that will offer us reasonably stable attractors as patterns to recognize in the phenomena in front of us.

I want to talk about this – the Internet, 2.0, and all of that. Part of what I say is that the Internet itself illustrates a sound set of these principles. The Internet itself is sort of a network that at least allows us to experiment with the sorts of patterns that lead to reliable knowledge and the sorts of patterns that don’t. And we’ve had the opportunity now for a number of years to study this. Jumping and leaping across a wider set of inferences here, but I’ve argued elsewhere that the types of networks, the networks that we would class as reliable or effective, instantiate certain properties.

They’re decentralized, so they don’t actually resemble that star cluster format that people like Clay Shirky characterize. They don’t create these power laws. They’re distributed; they’re not all in one place. They’re disaggregated; they’re composed of separate components. They’re disintermediated; things can communicate directly with each other. They’re dynamic, they change and we change; in other words, they’re plastic the way the human brain is plastic, allowing connections to form and to un-form. They’re disintegrated; it’s not all in one single system, but in fact, it’s in various subsystems. Jerry Fodor captures this in a book that I mostly disagree with, called The Modularity of Mind. They are, if you will, democratic.

The democrat principle is a special principle; I call this the semantic condition. Networks that follow these properties are network that are going to give us more reliable information, more reliable knowledge, patterns we can depend on rather than patterns that lead us astray. These reliable networks are networks that support or that instantiate autonomy. Each entity in the network functions on its own; it makes its own decision, it does its own thing. Diversity – the networks are not made up of the same kind of thing. The entities don’t all have the same properties, they don’t all have the same values. Openness – the network admits external phenomena that will help shape it. Connectivity – the entities in the network are, in fact, connected to each other.

Kathy Shields has nailed it: Not all patterns are valuable. Some of them are; some of them aren’t. And so what I’m describing here is the way of producing patterns that is reliable as opposed to producing patterns that are not reliable.

So now we come full circle. How is this practical? People are talking we have the old theory of knowledge and rules and all that, but it’s practical, it works. This is practical. How is it practical? Go back to the example I raised at the beginning of the talk. To teach the concept of brackets, would you use the same example over and over? Why not? Now we know – because of the need for diversity. Diversity more reliably produces patterns that can be trusted. Diverse experiences create better networks than monotonous experiences. So what you have here now is a set of principles. This set of principles you can apply to your current practices, which you do in the classroom, the type of software that you use, the way you organize a discussion group or a conference or a company, and these principles if followed produce more reliably effective companies or learning or classrooms or whatever. So for any one of these things, you look at the example, you look at the proposal, and you ask, is this distributed? Yes, that’s better. No, that’s not so good. So now you have, if you will, a rubric or a metric against which you can evaluate existing practice and proposed future practice. And that in a nutshell is the proposition I wanted to put before you today, so I thank you.

I think what’s changed is, especially for those of us who are the people in this conference, we’ve got this really big huge network that we can all see clearly for the first time. We never had networks like this before, certainly not networks we could get at. Now you go back to some of the early experimentations in social networks that are described by Watts and ? and others, how did they test the six degrees phenomenon? The six degrees phenomenon is any person in society is six steps away or six connections away, from any other person in society. Well, what they did is they mailed out letters and asked people to send letters to their friends to try to get the letter back, via the connection between friends, to the original source, and actually a whole bunch of those letters got lost and nobody knows what happened to them, but some of them made it back. This is the technology they were using – letters! You can’t study networks using letters, you can’t study networks using the postal system, it’s just not reliable enough, it’s not fast enough, you can’t grasp it. We have computational power now, we can actually create networks that we couldn’t create before. We can actually simulate connectionist systems, simulate patterns of interactivity between neural networks that we never had before. So what we have are ways of seeing that we didn’t have before, and because we have these new ways of seeing, that allows us to look at the same phenomenon, but from a different perspective. It allows us to use the metaphor that I raised at the beginning of the paper. It allows us to tilt our head in a certain way and see what we’ve always seen from a different point of view, and seeing it from this different point of view allows us to detect different phenomena in the same perceptual experience.

Questions and responses

I’m not sure exactly what you mean by abundance. Do you mean density and connectedness?

That’s an interesting question, because it depends on how you look at it. The learners, the individuals we have in society that are acquiring this information are the same physically as they were before – they have the same sorts of eyes, the same sorts of ears – and therefore, the amount of information that is coming into a person in 2007 is pretty much exactly the same amount of information that’s coming in in 1907. You see, you hear, so the question is more along the lines of where this information is coming from, the type of network that is producing this sort of information. If you look at the structure of networks through history, you’ll find that the connectedness is increasing, but also, too, the shape or structure of the network is changing. We come back to these power laws. In 1907 we had pretty extreme power law, where a few people would have influence over a great many people, and the great many people would have influence over very few. The classic example of that, we saw that come before in things like broadcasting, where you could have an Orson Welles or somebody broadcasting to 20 million people, the centre of a very large star. Over time, these big spikes are getting smaller. If you look at the A-listers on the Internet today, and these A-listers are talking to audience of what, ten thousand people, twenty thousand people; in very rare instances will they get up to the millions of people and then only for instances. So what we are seeing is a change in the nature of the connections. We’re seeing an increasing, if you will, modularity. We’re seeing an increasing focus on smaller but more dense sets of connections among – I don’t want to say groups, because I don’t mean groups – but among collections of individuals. We’re seeing more of an equal distribution of connections from person to person. To put it in slogan form, our networks are becoming more democratic, and hence more reliable. So it’s not a question of the amount of information; the amount of information is, to my mind, almost irrelevant. It’s the way this information is organized, the way this information is transported, the way we set up the system to pass this information from one person to the next. And as the influence of any given source decreases, the reliability of our network increases. There are so many historical examples of that which I won’t get into here.

That’s the end of our recording. They’re going to talk about Second Life now.