Large-scale Thalamocortical Model

A fellow student in my lab is basing some of his preliminary work on neuron models devised by Eugene Izhikevich, who published a book last year in which he described a system for modeling the diverse spiking behavior of many types of neurons with an elegant set of equations. In a paper co-authored with Gerald Edelman (of Neural Darwinism fame) they implement a model of the cortex and thalamus and their interconnectivity.

Here are some features of the model:

1) Simulates one million multi-compartmental neuron models of 22 basic types
2) Includes approximately half a billion synapses
3) Macroscopic connectivity is based on data derived from diffusion tensor imaging (DTI) of magnetic resonance image (MRI) scans of the thalamus and cortex
4) Microscopic connectivity is based on reconstruction studies of cat visual cortex
5) Synapses are modified through spike-timing dependent plasticity (STDP)

Now, the thalamus is the part of your brain through which almost all sensory input is routed before being sent to the cortex (the exception is olfactory input). This structure is about the size of the end of your thumb, and it is a place through which nearly all of your input from the world flows (visual, auditory, and tactile). But like most brain areas, exactly what it does is not very well understood. For example, we know that it does not function merely as a relay station. There is extensive feedback from the cortex back to the thalamus, creating a thalamocortical loop. Why would the cortex need to send information back to the thalamus if it's just a relay station? Some theorists have proposed that the thalamus is something like an active blackboard, maintaining a constantly updating sketch of the world. Others have proposed that the loop is a way of keeping recent events in a kind of short-term buffer.

Whatever the case, the Izhikevich and Edelman model does not simulate the stream of sensory input from the world. So how does anything happen in the model? Well, initially it is quiescent. The modelers get it going by causing random neurons to spike, which effectively jump-starts ripples of activity throughout the system.

One interesting finding is that the brain state is very sensitive, so much so that the alteration of the spiking activity of a single neuron radically alters the global firing patterns throughout the model within less than half a second. This seems a bit counterintuitive. We might expect that the brain is robust to small changes. After all, neurons can be fairly noisy (that is, they don't always fire reliably), and they also tend to die off. So either their model is overly sensitive to small perturbations (i.e. the butterfly effect) or this really is a reflection of the sensitivity of real neural systems. Either way, it's an interesting result.

One last comment...the authors state that they "started with the thalamocortical system because it is necessary for human consciousness." In discussing the paper with another student, I mused about the ethical ramifications of this kind of simulation. I seriously doubt that a simulation of a million neurons evoked anything like consciousness when randomly jump-started, but then, consciousness is a very poorly-understood phenomenon. I told the other student that I was reminded of Johnny Got His Gun, an anti-war novel in which a soldier is wounded such that he loses all senses but touch, all his limbs, and most of his face. He tries to communicate by banging out Morse code with his head on the hospital bed.

"But this thing doesn't even have a head to try to bang out Morse code," I joked, before I realized just how creepy that sounded. So like I said, I seriously doubt we have to worry about the ethics of simulating consciousness at this point. There are about 100 billion neurons in a human brain and about 20 billion in the cortex, while this model uses one million neurons. But it's something to at least ponder, and definitely something to consider more as models become more and more sophisticated.

E. M. Izhikevich, G. M. Edelman (2008). Large-scale model of mammalian thalamocortical systems Proceedings of the National Academy of Sciences, 105 (9), 3593-3598 DOI: 10.1073/pnas.0712231105

Why Do You Want to Build a Superintelligent Artifact?

One thing I'd really like to see is a survey of researchers working in artificial intelligence or closely related disciplines asking what their motivations are. Nearly everyone I've talked to who works in AI dreams of ultimately building a human-level or higher intelligence. But why?



Most of the super-intelligent machines from fiction and the movies (The Matrix, 2001: A Space Odyssey, Terminator, and on and on) don't tend to have humans' interest at heart.



Or do they want to create something more like Data?



My guess is that most AI researchers are technological optimists, and assume that whatever they happen to engineer will be benevolent. However, history teaches us that how a given technology is used depends on the character of the culture that is wielding it. I also wonder how honest researchers would be, even in an anonymous survey. Another probable result would be that many researchers simply aren't conscious of their motivations. I tend to wonder the extent to which technological innovation advances simply because people want to make cool stuff. Money is of course another motivation, but while the rewards of manufacturing androids would be obvious, such long-term goals are highly speculative, and most bright people could earn a lot more money by focusing on more conservative approaches.

Personally, I'm optimistic that human-level AI will eventually happen, but I'm doubtful that it will happen any time soon. I think the path will involve machines that, like humans, learn most of what they know, rather than having it innately programmed. This will mean a very long training period, and also entails that the character of the machines will be closely related to the type of training they receive. Anyway, I'm much more interested in building a Data than a Skynet. But ultimately our relationship to whatever we create is contingent upon how much we think and plan about the consequences of what we're working on.

Methods for Studying Minds

Cognitive Science suffers from the peculiar problem that we are using our cognition to try to understand cognition. Science is about reducing bias and subjectivity through peer review and reproducibility, but the mind is the source of subjectivity, so how in the heck do we study it.

As I see it, there are four broad methods for studying cognition:

1) Introspection

This is probably the least reliable because it is the most subjective, but I don't think it should be discounted out of hand for that reason. A researcher might gain insights into their own cognitive functions by reflecting on how they might work and paying close attention to how they think, the ways in which their memory works, etc. Obviously, the main drawback is that an individual has privileged access to this kind of information, so it is not subject to unbiased reproducibility.

2) Verbal Communication

This method basically entails asking people what their states of mind are. If I ask you whether you dream, and you say "yes," this is increased evidence that humans experience dreams. The more specific the information gets, the less reliable it becomes, because it relies on relaying information derived from introspection to another party. Still, it is a necessary inclusion in the toolbox.

3) Observation of Overt Behavior

This is the most popular method of studying cognition, and for a time it was the only one. It involves measuring aspects of the outward behavior of an individual, such as reaction time or accuracy while performing a verbal or mathematical task. For a non-human subject, measurements might be made regarding successes or failures at solving a particular problem, such as pressing a series of buttons to receive food. The main problem with relying solely on this method is that the cognitive system is a black box. We only look at the inputs and outputs, and try to make inferences about how the processing might be going on, which is kind of like trying to figure out how a radio works just by listening to it, and not ever looking inside. Which brings us to the last type of methodology...

4) Observation of Cognitive Mechanisms

In the case of biological organisms, this primarily involves neuroimaging, such as EEG, PET, fMRI, etc. Rather than measuring aspects of overt behavior, we look at the mechanisms underlying the processing of information, and how they are behaving. If a monkey is given a task to discriminate between a cube and a sphere, we can directly measure the activity of the cells in their nervous system, or blood flow to a particular area of their brain, which might give some insights into how the task is processed. If we're studying how an artificial intelligence solves a particular task, we have direct access to the algorithm that accomplished the task, and should be able to determine how it did what it did.


When it comes to studying humans, we have all four methods at our disposable, to greater or lesser degrees. Ethical considerations place limitations on the type of data we can gather using these methodologies. For example, lesioning studies, where part of the brain is surgically removed, with monkeys are quite common. These simply aren't ethical with humans. But we do have cases where people suffer damage to brain areas due to injury or disease, which do let us carry out these kinds of studies in an indirect way.

With non-human subjects, such as non-human animals or AIs, we can only use #3 and #4 (at least until those subjects can communicate sufficiently in a natural language). Sign language taught to chimpanzees and gorillas isn't sufficiently rich to communicate meaningful information about their states of mind, and no current AI is close to being able to communicate in a natural language.

Data from all these sources continues to pile up, and neuroimaging especially has advanced as a technique for measuring the behavior of cognitive machinery, but we're still at a loss for strong theories in which to incorporate all the information.

Is Consciousness Special?

I've had discussions with people before who make the claim that consciousness is somehow going to be forever beyond the realm of science. Here's a Bloggingheads.tv discussion between Eliezer Yudkowsky and Jaron Lanier. About 25 minutes in, this is exactly the kind of point that Lanier tries to make:



If I'm understanding his point, he's saying that he doesn't understand consciousness, but that an understanding of consciousness will not be possible under the assumption that it is a product of physical things (like neurons) carrying out their function in the context of a physical system (like the human brain).

Yudkowsky calls him on this, by basically asking how you can make claims about how something is or isn't going to be understood if you don't understand it. It's a great question, but Lanier does what he does throughout the discussion, which is either laugh and move on to another point or make some kind of ad hominem slur against Yudkowsky by calling his adherence to science a kind of religion.

Lanier ironically states that making such groundless, pessimistic claims somehow makes him a better scientist. Huh?

Look, everything in the world was mysterious before it was explained. Lightning and thunder, for example, were very mysterious. Ancient people came up with initial explanations having to do with the actions of supernatural beings. Those explanations turned out not to be very good. We actually started to get somewhere when someone went, "Hey, wait a minute...maybe there's a reasonable explanation for what's going on here that we can understand."

That assumption, and not a hypothesis or theory or experiment, is the beginning of science and the first step on the road to knowledge. If you don't take that step, and automatically assume that either the explanation is supernatural and/or that you will never be able to understand it, then what you have done is guarantee that you will never understand it because you've given up before you even tried.

The history of science is one in which we have made progress in our understanding of the world by making the default assumption that things are the result of natural processes that work in orderly ways according to principles that we can figure out if we work hard enough.

It may very well be the case that there are hard limits to human understanding, and that there are things we are not going to be able to figure out. The nature of the universe may be one such question, whether or not there is a single frame of reference or some kind of multiverse, why the universe as we know it is expanding and whether or not it is a single case of such an expansion or one of many cycles of expansion and collapse. A grand unified theory of physics that reconciles quantum mechanics and relativity may be beyond our understanding. And so might consciousness. But the fact is, we have barely even begun to try to systematically understand these things. A few rare people have pondered such questions for millennia, but organized science as an institution has only really picked up speed in the last 150 years or so, which is really a very small amount of time.

So isn't it just a tad early to throw in the towel?

Are Brains Digital or Analog?

Last year Chris Chatham wrote up a great post entitled 10 Important Differences Between Brains and Computers. There's a wealth of topics to discuss in reference to the post, but I want to focus on what he lists as the #1 difference:

Brains are analogue; computers are digital
It's easy to think that neurons are essentially binary, given that they fire an action potential if they reach a certain threshold, and otherwise do not fire. This superficial similarity to digital "1's and 0's" belies a wide variety of continuous and non-linear processes that directly influence neuronal processing.

For example, one of the primary mechanisms of information transmission appears to be the rate at which neurons fire - an essentially continuous variable.

There's more, but this is the essential part. What's interesting is that Chris points out that one way in which information is conveyed in the brain is by the rate of fire of neurons. But then he ignores the fact that we know that information is carried by the timing of individual spikes, and he categorically labels the brain as "analogue."

A nice metaphor for neurons is a leaky bucket. When they are receiving incoming activity from other neurons, you can think of that as water trickling into the bucket. This is analogous to a charge building up on the cell membrane. But the membrane has resistance, so it is "leaky", which means in the bucket analogy that there are is also a tiny hole in the bottom of the bucket. If the water you're putting in doesn't exceed the leakage, then the level of water will never rise. Once the water reaches a particular level, a threshold, then you can think of the bucket being tipped over and sending all its water to all the other buckets it connects to. This is analogous to the firing of a neuron. It then resets until it is filled back up to its threshold level.

So before it reaches threshold, a neuron functions in an analog fashion. When it reaches threshold, it generates an action potential, or spike, which is a binary signal. But then, as Chris points out, if we count the number of spikes within a given time frame, that rate of fire can be measured in an analog fashion.

In my own work, I used to use artificial neurons that modeled only the average rate of fire of neurons. These are known as rate-coding neuron models, and a very common function that approximates the firing rate is the sigmoid:

But if you use such a model, you're assuming that no information is being carried by the timing of individual spikes, because you're averaging that information away.

But we know of particular examples in which information is conveyed by individual spikes. One very famous and interesting example is the auditory system of the barn owl. See "Hebbian learning of pulse timing in the barn owl auditory system" by Wulfram Gerstner, Richard Kempter, J. Leo Van Hemmen, Hermann Wagner for a great overview.


Basically, when a mouse makes a sound, the sound waves reach each ear of the barn owl at different times, because the ears are spaced apart. The owls auditory system is able to determine where the sound came from by comparing the relative timing of the sound reaching each ear, and this information is learned and conveyed via the timing of individual spikes. By the way, the image isn't of a barn owl shooting a mouse with laser vision (though that would be cool). It's meant to show how the sound of the mouse squeak reaches each of the barn owl's ears at slightly different times.

We also know that an important aspect of learning throughout the brain involves the relative timing of individual spikes. If a neuron (A) fires just before the neuron (B) it is connected downstream to, then the synapse will "strengthened", or modified in such a way that the next time neuron A fires, it will be more likely to cause neuron B to fire:

However, if the order of firing is reversed, then the synapse is "weakened":

The strengthening and weakening of synapses in this way is known as spike-timing dependent plasticity, or STDP. While there are a number of other ways in which synapses are modified in the brain, these particular mechanisms are thought to underlie many important aspects of learning, and they should not be ignored.

So, the answer to the question "Are Brains Digital or Analog?" is a perhaps unsatisfying "both". Some of the ways in which neurons communicate and undergo modification via learning are based purely on all-or-nothing signals in a digital way. In other cases, information is conveyed by the rate of fire of neurons in an analog manner.

But then, computers emulate analog functions as well, so they are neither distinctly digital or analog. In fact, the dichotomy turns out not to be all that sharp in many domains. What's important is to know in what ways the brain is analog and in what ways it is digital. Both will likely figure into any coherent explanation of how the brain works.

For my own part, I've begun working with spiking neuron models, specifically what are known as leaky integrate-and-fire models. I've become increasingly convinced about the importance of the role of time in understanding cognitive processes, and spiking models allow for communicating information both by the timing of individual spikes and by their rate of fire, while rate-coding models only allow for communication via average firing rates. That's not to say that rate-coding models don't have a lot to teach us about certain aspects of cognition, just that they are limited in their ability to do so.

Invariant Representations

I just finished rereading Jeff Hawkins' On Intelligence on audiobook. One of the keys to developing intelligent systems is to enable the system to learn invariant representations of things in the world, and then use current information to make predictions about what's coming next.

An invariant representation is a way of storing information so that if the information appears in a slightly different form, it is still recognizable. For example, you recognize the melody "Happy Birthday" no matter what key it's played in. What let's you recognize it are the intervals.

One example Hawkins uses is the "train station example". Let's say you live in a town a long time ago, and your sweetheart is supposed to come to the town to live with you, and they're arriving by train. Every day you go to the train station, but your sweetheart doesn't arrive. You know that two trains run per day, and you've gotten a letter saying they'll be on the later train. After a couple of weeks of visiting the station, you see a pattern. The morning train arrives at different times, but the afternoon train is always exactly 4 hours later than the morning train. You develop an invariant representation of the train schedule. So if the morning train arrives at 10:14, you know that the afternoon train will arrive at exactly 2:14. If the morning train arrives at 9:27, you know the afternoon train will arrive at exactly 1:27. What you have encoded is relative information, rather than absolute. So given your representation, and the time of the morning train for that day, you can reliably predict when the afternoon train will arrive.

Same thing with vision or audition. If you have an invariant representation of an object, like a dog, it doesn't matter if the lighting conditions are slightly different, or that the dog is near you or far away, or that it's upside down or rotated. Given you invariant representation and information about, say, where the ears are, you can predict where the eyes, legs, and tail are going to be.

This reminded me of the difference between raster and vector graphics.



Raster graphics are bit-for-bit encodings of images. They encode absolute information, about every single bit. Sometimes this is good, but sometimes it's bad, as when you want the image to scale without losing resolution.

Vector graphics, on the other hand, store an invariant representation of the image, and the computer renders objects given certain information.

For example, a circle stored in raster graphics would encode the position and color of every single pixel that makes up the circle. When you zoomed in, the circle would look "blocky". And you'd need a lot of information to store the image.

A circle stored in vector graphics would need to know the formula for the circle, the radius, and other relative information. Then all it would need is the center, so it would know where to render it.

Hawkins makes the claim that AI researchers have not used such an approach, but rely on absolute encoding in order to teach machines to recognize patterns and produce actions, but it would surprise me if the general approach had not been taken.

How Do We Learn What Things Are?

You probably don't think about it as you go about your daily routine, but you never see the same image twice. That is, the same pattern of light, shadow, and color never fall on your retina in the same way. When you see a friend's face, and you're talking to them, you have the subjective perception of stability and constancy. In reality, your eyes are flitting all over the place. When your eye makes a rapid movement to another location, that's called a saccade. When it sits still for a very short time, that's called a fixation. The average person makes three saccades per second.

Here's a famous image from a study by Yarbus from 1967:


It shows a picture of a female's face, and the eye tracking pattern of someone who looked at the face. When we see a face, we tend to flick our eyes constantly across it, though not in a random way. You can see that we tend to concentrate on prominent features like the eyes, nose, and mouth, which are rich in information that help us distinguish the face from others. You don't spend a lot of time looking at cheeks, because they don't give you as much information as the eyes do.

And while your eyes are constantly moving, most interesting objects, like people, are also in motion. So is your head and body. The successive images on your retina change very rapidly all the time, and yet when you stand in front of a painting in a gallery, it doesn't feel like you're viewing a movie with cuts every third of a second. But that's what's really going on.

So if what we're seeing is a rapidly changing movie, even when we're looking at a stationary object like an apple, how the heck do we learn that all these rapidly changing images relate to the same object?

One idea is that our brains learn to cluster together things that appear close together in space and close together in time. The idea is that, even though your eyes are flitting all over the apple, your brain is keeping track of how far your eyes are moving. If that's a small region, and the successive images occur close together in time, it's a good bet that the thing you're looking at is a coherent entity.

Same for images that vary even more dramatically over time. Think of something like an elephant. You recognize it from far away and up close, even though the image falling on your retina may be many times smaller or larger. You recognize it when it's rotated, even though it may look very different from the front and from the back. You also recognize it when it's upside down. You recognize it if it's painted green, on a cloudy day or a sunny day. Computers fail miserably at learning the recognize objects with this degree of variance. So how the heck do you do it?

Dileep George calls the method by which we group successive inputs into one representation temporal pooling. The paper by Cox et al. that I'll be referring to here discusses the temporal contiguity of input as a way of grouping it together. The idea is simple: Things that occur close together in time are probably closely related, e.g. they're a unitary object or concept.

Cox et al. did a clever experiment. They used artificial objects shown here:


The objects are similar, but they are distinct. What they did was, they had subjects look at a cross in the middle of a computer screen. Then they showed an object, like object A, to either the right or left of the cross, in the subject's peripheral vision. The subject would naturally saccade to the object. Here's a figure showing what a typical trial would be like:


Now, in another condition, the first three steps would be the same. Subject looks at the cross, object A appears to one side, and the subject saccades to fixate on the object. But in this condition, in the very short time it took the subject to saccade, they switched the object from A to A'. When you made a saccade, you're temporarily blind, so the subject is not even aware that the objects have been switched. Here's what that kind of trial looks like:


After a number of such trials, they tested the subjects with a "same/different" paradigm. This means they basically showed them A and A' together and said, are these the same objects, or are they different objects? Subjects from the normal condition found it easier to distinguish between A and A', while the subjects from the swapped condition were more likely to say that A and A' were the same object.

So what does this suggest? That we're more likely to group together images that occur in rapid succession. This type of work is very closely related to the modeling I'm doing for my dissertation work, and I'm going even further and suggesting that that clustering input based on spatial and temporal contiguity is the fundamental mechanism by which we learn...not just visual objects, but music, language, tactile input, and on and on.

Cox et al.'s experiment is a very clever, very nice way of demonstrating how the principle works in the visual domain.

David D Cox, Philip Meier, Nadja Oertelt, James J DiCarlo (2005). 'Breaking' position-invariant object recognition Nature Neuroscience, 8 (9), 1145-1147 DOI: 10.1038/nn1519

What is Hiearchy?

In my primate cognition class last semester, we were having a discussion about baboon social structure, I think, and we were talking about it being hierarchical. The professor asked me what "hierarchical" meant. I was taken aback for a few moments, and couldn't come up with a response. It's good to question the meaning of words you tend to take for granted.

After reading On Intelligence in 2005, Hawkins convinced me that time and hierarchy are crucial to understanding cognition. Whether his company's particular implementation is the right technological path is an open question, but I think in terms of identifying the important theoretical concepts, he's dead on.

So what is a hierarchy?

In The Sciences of the Artificial, Herbert Simon says:

By a hierarchical system, or hierarchy, I mean a system that is composed of interrelated subsystems, each of the latter being, in turn, hierarchic in structure until we reach some lowest level of elementary subsystem. (emphasis mine)

I didn't find this definition very satisfying, though it's all right.

Wikipedia does a nice job, I think:

A hierarchy is an arrangement of objects, people, elements, values, grades, orders, classes, etc., in a ranked or graduated series.

...

Items in a hierarchy are typically thought of as being "above," "below," or "at the same level as" one another.

Not bad, though I'd truncate it down to simply:

A system of ranked elements.

Rank, it seems to me, is the most important aspect of hierarchy, and this simple working definition seems to capture nicely the essentials of hierarchy.

So why is hierarchy important to understanding cognition?

Well for starters, because the world is hierarchical. The universe is composed of galaxies, which are composed of subsystems, including solar systems. All matter is composed of atoms and molecules, which are composed of subatomic particles. And multicellular organisms, including ourselves, are composed of subsystems (nervous, circulatory, etc.) that are composed of organs composed of cells composed of molecules and so on.

And that's just spatially. Temporally, our lives are composed of stages (early childhood, adolescence, the college years, etc.), which are hierarchically divided down to individual moments.

Our neocortex is hierarchically-arranged, and I believe its architecture has evolved specifically to exploit and encode the hierarchical nature of the world around us. In The Quest for Consciousness, Christof Koch refers to the neocortex as "quasi-hierarchical," by which he means that it is not a strict hierarchy. Some cortical areas don't report directly to the area just above them. Many connections skip levels, like a private reporting not only to his sargeant, but to the general as well.

I don't really like the term "quasi-hierarchy" much, though. The neocortex is no less hierarchical because of this arrangement. If a system is composed of elements which are all the same rank, it's not a hierarchy. If there is any difference in the ranking of elements in a system, then it's a hierarchy. I'd prefer qualifiers like "strict" and "flexible" to describe the extent to which the relationship between elements in a hierarchy is between adjacent levels.

And I'll leave you to ponder this adapted diagram of the organization of the macaque monkey's visual from Felleman and Van Essen (1991). The information flows from bottom to top. The bottom part is information coming from the eyes, and the hippocampus sits on top. Our visual processing system is very probably arranged in the same way.

CogSci 08: Day One

I missed the plenary talk this morning, mostly because they've scheduled all the plenary talks for the wee hours of the morning, and this one started at 8:15am.

So the first session I attended was a symposium of related talks on conceptual primatives. Jerome Feldman was the chair, and he first said a few opening remarks to frame the other talks. He talked about how the symposium was in the context of Unified Cognitive Science, which means that evidence from multiple subfields would be needed to address the question of the nature of conceptual primitives.

He gave a URL for a working list of possible conceptual primitives, but I wasn't able to jot it down. It is probably in the proceedings, so I'll look for it later. Among the examples were things like:

day/night
near/far
time
animate/inanimate
person
gender
age
possession
event/thing
number
goal
support
containment
true/false

He said that they were advance a "theory neutral" approach, and the only question asked of him was to clarify what he meant by "theory neutral". He said he couldn't clarify it, because it was a self-explanatory term. Fair enough, but I was still wondering what he meant by "conceptual primitive". I assume he means a concept whose representation cannot be broken down any further and represented as a combination of other representations. But who knows? I can definitely see how some of the concepts listed above could be represented as combinations of other concepts. If this isn't the meaning of "primitive", then what does it mean?

Anyway, the first actual speaker was Leonard Talmy. He started out by saying that every language has two subsystems:

open-class - morpheme classes that are large and easy to augment (e.g. the roots of nouns, verbs, and adjectives)
closed-class - morpheme classes that are small and difficult to augment (e.g. bound inflections and derivations, and free prepositions, conjunctions, and determiners)

He said his talk would focus specifically on closed-class forms that relate to representing space. He talked about a methodology of choosing a figure and ground and a particular closed-class form and determining what the constraints are for using that term. The extended example he gave was board (figure) laying across (the closed-class spatial term) a road (ground). He gave several examples of situations where it is entirely appropriate to say "The board lay across the road" and situations where it is not.

For example, if the board is actually perpendicular to the road, it wouldn't be acceptable to say it is laying across it, but something like "The board is sticking out of the road." But he also said a constraint was that the board is touching both sides of the road, in which case you would say "The board lay over one side of the road." But I wouldn't have a problem saying a board lay across a road as long as it almost reached the other side, even if it didn't touch it.


He also said the axis of the figure needed to be horizontal, such as "The spear hung across the wall" vs. "The spear hung up and down on the wall." But that doesn't seem to hold for his own example of the board and the road. Both of these work for across:

But I get the basic idea that there are constraints on the usage of particular terms...that's what makes them distinct concepts. I'm not sure how this helps us narrow down conceptual primitives, whatever those are.

I need to get to the next session, so I'll wrap this up for now. There were three more speakers, and I'll have a separate entry for some or all of what they covered as well.

Home Again, Then Out Again

The New Orleans trip was fun. I did indeed shove a couple of beignets in my face.

But I'm already heading back over there tomorrow. We're flying out of New Orleans, heading for Washington DC and the Cognitive Science Society Annual Meeting. I really enjoyed the conference last year. True to the spirit of cognitive science, there really is a wide range of theoretical and experimental work displayed relating to cognition. The problem is, as with most conferences, you can't see everything you want to.

But I'll blog about the stuff I am able to see and take decent notes on.

Naturalistic Dualism

In one of the coffee breaks between sessions at GECCO, I was introduced to the concept of naturalistic dualism, which is a stance on the nature of consciousness. The term was coined by David Chalmers, and is described here.

If I understand the basic idea, naturalistic dualism holds that consciousness:

• Is not a reducible phenomenon
  
• Cannot be explained in terms of function
  
• Is fundamentally different from anything physical or any function of anything physical, and is therefore a qualitatively different entity from anything else in the known universe
  

If anybody knows more about this philosophical stance and I'm screwing it up, please correct me in the comments.

Here are some sections from Chalmers paper. See if this makes sense:

Purely physical explanation is well-suited to the explanation of physical structures, explaining macroscopic structures in terms of detailed microstructural constituents; and it provides a satisfying explanation of the performance of functions, accounting for these functions in terms of the physical mechanisms that perform them. This is because a physical account can entail the facts about structures and functions: once the internal details of the physical account are given, the structural and functional properties fall out as an automatic consequence. But the structure and dynamics of physical processes yield only more structure and dynamics, so structures and functions are all we can expect these processes to explain. The facts about experience cannot be an automatic consequence of any physical account, as it is conceptually coherent that any given process could exist without experience. Experience may arise from the physical, but it is not entailed by the physical.

I don't get how it is "conceptually coherent" that any given process could exist without experience. The process that gives rise to experience cannot exist without giving rise to experience.

Look, we can't explain how consciousness works. But every single phenomenon that we now understand in physical terms was once thought to be some utterly mysterious, qualitatively different stuff, usually couched in supernatural terms. The history of science is the account of the majority of people thinking that phenomenon from lightning bolts to disease was some mysterious result of supernatural workings. A few brave, skeptical people thought, "Wait a minute...maybe there's a reasonable explanation for that." And then they did the hard work to try to figure it out.

Now Chalmers has a response to this:

It is tempting to note that all sorts of puzzling phenomena have eventually turned out to be explainable in physical terms. But each of these were problems about the observable behavior of physical objects, coming down to problems in the explanation of structures and functions. Because of this, these phenomena have always been the kind of thing that a physical account might explain, even if at some points there have been good reasons to suspect that no such explanation would be forthcoming. The tempting induction from these cases fails in the case of consciousness, which is not a problem about physical structures and functions. The problem of consciousness is puzzling in an entirely different way. An analysis of the problem shows us that conscious experience is just not the kind of thing that a wholly reductive account could succeed in explaining.

This is if you buy his argument that conscious experience is not the kind of thing that can be explained in terms of physical structures and their function. Which I don't.

This seems to me to be a very counterproductive agenda, especially to a scientist, and especially to a scientist bent on trying to figure out how the mind works.

Poker: Human vs. AI

Having defeated humans at chess, AI programmers are looking for other game domains that require more human-like skills. Good candidates are Go, Bridge, and Poker.

Right now there's a human vs. computer poker competition taking place:

Developed by an artificial intelligence group at the University of Alberta in Canada, Polaris will be pitted against several professionals at the Rio Hotel between July 3rd and 6th. Its human opponents will include Stoxpoker.com coaches Nick Grundzien and Ijay Palansky along with Matt Hawrilenko, all of whom have well over $1 million in lifetime winnings from playing poker.

They tried this last year, and the program did reasonably well, but didn't win. The reason poker might be a better benchmark for AI is that in involves making decisions with incomplete information (e.g., your opponents hands). It also requires probabilistic reasoning (unlike chess, which is completely deterministic). Also, it has been shown that any optimal strategy in poker requires some bluffing, which makes sense. If you can gain an advantage by misrepresenting your strength (either under- or over-representing) you're going to lead opponents into giving you more of their chips. But knowing when and how to bluff is difficult.

So it's an interesting story, but I thought this quotation near the end was pretty dumb:

"It's possible, given enough computing power, for computers to play 'perfectly,' where over a long enough match, the program cannot lose money," said associate professor Michael Bowling. "Humans will always make some mistakes, meaning the program will have an advantage."

I ran into this same fallacy when I read a paper about Tic-Tac-Toe several years back, in which they argued that a program that never loses at Tic-Tac-Toe is "playing optimally". Well, no. If you want to define it that way, good for you, but it's a very poor definition.

To play "optimally" or "perfectly" doesn't just mean that you avoid losing, but that you maximize wins against weaker opponents. I don't think we'd call a poker player that never lost money but just barely made money a "perfect" player.

Computing Beyond Turing: Jeff Hawkins Talk at UCSD

As you may or may not know, I'm a big fan of Jeff Hawkins, entrepreneur, engineer, and cognitive theorist. You really should read his book On Intelligence, if you haven't already. It's changed the way I think about cognition and has been a driving force for the direction of my dissertation.

There are some older talks out on the web of Hawkins explaining the theory that's driving the technology he's now working on, but here's a fairly recent video of a talk he gave at UCSD this year:




He does review the theory, but he also gives a demo of the NuPIC software and describes some of the recent additions they've made in the past year. It's about 1:15.

One of the things that's funny is that the title of the talk is "Computing Beyond Turing", and Hawkins spends some time talking about Turing and the Universal Turing Machine. When a questioner points out at the end of the talk that he didn't really talk about computing beyond Turing, he admits it, so it's really not a good title for the talk. The hierarchical computation Hawkins is talking about is still carried out on a Turing machine. If anything, it's a subset of Turing computation.

Still, it's an interesting and worthwhile talk if you care to invest the time in it.