Recent Advances in Cognitive Systems

Roland Piquepaille writes "ERCIM News is a quarterly publication from the European Research Consortium for Informatics and Mathematics. The April 2003 issue is dedicated to cognitive systems. It contains no less than 21 articles which are all available online. In this column, you'll find a summary of the introduction and what are the possible applications of these cognitive systems. There's also a picture of the cover, a little robot with a very nice looking blue wig. And in A Gallery of Cognitive Systems, you'll find a selection of stories, including links, abstracts and illustrations (the whole page weighs 217 KB). There are very good pictures of autonomous soccer robots, swarm bots, cognitive vision systems, and more."

Comments

Re:Cognitive Science

[Max Romantschuk]

Noun ; 1. The current scientist scam, which has replaced the older artificial intelligence scam with its more robust resistance to criticism and even more byzantine theories.

Actually, cognitive science does not replace AI. The goal of cognitive science is to figure out how our brain works on a functional level. Where neurology studies the actual chemical reactions and neural activity, cognitive science studies how the "hardware" works to achieve our thought processes.

One good example is how the brain works out an image of the mismash of neural impulses going through the retinal nerves. The resolution of the eye is actually quite low, and the "pixels" aren't ordered in any linear fashion. The brain does an enormous amount of processing to form an actual image. This is why babies can't see, even though the optics work. The brain needs to develop the processing algorithms in order to make sense of all the information coming in.

Of course, all of this is theory, and subject to scientific dispute :)

You can teach a computer to think

[ObviousGuy]

But doing so doesn't relieve you of your responsibility to think too.

The title reminds me of an article in AIR

[Jonathan]

The Annals of Improbable Research, the humor magazine for scientists, once had an article entitled "Advances in Artificial Intelligence". After the title and author affiliations, the page was appropriately completely blank...

Re:The title reminds me of an article in AIR

[Anonymous Coward]

The problem is, every time an advance is made in the field of AI, that advance is immediately redefined "not AI". Voice recognition, which you can now walk into a shop and buy software for, is now "not AI". Chess playing, "not AI".

Essentially, AI is used to mean "stuff computers can't do yet".

People say "but the computer's just doing maths". Well, that's the point, isn't it? It might be that an AI powerful enough to be mistaken for human is simply horrendously complex, not unattainable, needing the sum of all those little incremental advances that AI researchers keep making.

Actually,the thing that annoys me most is that people associate Lisp with 80s AI, when in fact modern Common Lisp is an excellent multiparadigm language for all sorts of problems, and a much better fit for large software systems than, say, Java.

Re:The title reminds me of an article in AIR

[alienmole]

The problem is actually largely self-inflicted by AI researchers, who have at various times used AI as a gee-whiz hook to justify all sort of research that are at best, peripheral applications of a weak form of AI. Even scientists pay a price for overuse of hype.

"Real" AI would emphasize the "intelligence" part and be capable of, for example, learning the rules of a new game or process from a natural language description and trial and error, and then being able to perform said process. Anything less is pretty much just dicking around with heuristics.

Anyone who ever claimed that machine vision or chess playing or voice recognition was AI, was either confused or guilty of the charge in the first paragraph above. Even before those things were first achieved, the people actually working on them had a pretty good idea of how they could be achieved without anything like what we normally consider intelligence - and they went on to prove it.

Re:The title reminds me of an article in AIR

[KingJoshi]

I think you slight the progress made, the difficulty of the task and overgeneralize on the AI community (on purposes and approaches of various people). You also have a clarity on "intelligence" which tends not to be so clear.

But I'd like to bring to your attention a research project going on at my school (Michigan State University) which I think is different from other "AI". I didn't see it mentioned from glancing the article.

The attempt to is create a robot that learns and develops as a baby would. A key point is that it develops its own representation of the world. I disagree on some issues with the professor, but I think he has the right general idea.

Here's a link to slides explaining the approach and another link to the main research page.

Re:The title reminds me of an article in AIR

[alienmole]

Thanks for the links.

I don't mean to slight the progress made, and I also didn't mean to criticize all AI researchers.

Perhaps a better way to describe what I was getting at is that there's an unfortunate feedback effect that happens with these advanced applications, where: researchers say things which excite the general public because they describe things that sound amazing and desirable; researchers notice said excitement and connect that with increased funding; researchers exploit excitement by attaching loaded buzzwords like "AI" to all sorts of vaguely related research projects. But what the public heard or believed initially, is never actually delivered, and what is delivered doesn't seem nearly as exciting as the original vision. If not for this effect, the first "AI crash" would never have happened.

Of course, what I've described is to some extent how the promition of just about any project or product works. The difference with advanced applications like AI is that the ultimate end goals - which are often brought up as justifications for the work - are so far from achievability that expectations are dashed much more than usual when the projects finally reach some kind of fruition - if they ever do. Much of the audience then feels as though it was burned, and could care less about the fact that "real" AI is so much harder than any other software that's been developed to date. They simply perceive that what was "promised" was not delivered.

Like public companies which learned to carefully manage their earnings so as to remain in line with Wall Street expectations, researchers in these fields need to be careful about expectations management if they're going to promote their projects publicly - unless they have something concrete they're going to be delivering in a finite and predictable timescale.

I do think a term like "Cognitive Systems" is much less likely to suffer from these kinds of problems. Many things which could reasonably be called cognitive systems research could not, without significant qualification, really be called artificial intelligence research.

Real world problems and neuroscience

[Neuronerd]

The great thing about the recent development in so-called cognitive systems is that they start to address more real problems. The time of toy problems is over. It is not enough to just follow a line. Only the challenge from the real world can make algorithms in any way "clever" or meaningful.

This is why I find it truly inspiring that so much research is going into these systems these days.

Sadly however most of neuroscience these days is still far from these questions. Most electrophysiologists that for example study the visual system show it trivial stimuli such as bars or gratings. In some sense a system can only show its capability when the stimuli are rich enough.

Nevertheless there is clearly a move these days towards larger more interesting problems even in neuroscience. We should be inspired by the works of the roboticists.

Re:Real world problems and neuroscience

[TomorrowPlusX]

Sadly however most of neuroscience these days is still far from these questions.

This is interesting to me, for several reasons. I'm working on robotics in my free time, mainly not cognitive stuff but lower level autonomous muscular control and feedback loop stuff. But anyway, my girlfriend's studying neuroscience and she, like many (too many) of her peers, finds absolutely NOTHING interesting in cognitive research.

All they care about is the mechanics (which is important) but I think they consider cognition to be a peculiar but unimportant side effect of the rest of the complex process.

So, as a fellow who's spent years writing code to try to do intelligent stuff, and more recently robots to carry these actions out, it's somewhat frustrating to be in a bar with a bunch of neuroscientists and hear them dismiss cognition as irrelevant.

Not all cognitive scientists do that.

[fireboy1919]

Actually, I'd say that not very many are doing that.

The goal of all the cognitive scientists I've met is to make machines think, just as with A.I. In fact, I've always heard, and was told in my AI class, that A.I. is a branch of cognitive science.

However, there are many approaches to machine thinking that are not considered part of A.I.:
neural networks, SVMs, computer vision (signal interpretation), modeling.

So what does A.I. cover then? Well, it's not exactly well defined. If you read A.I. textbooks, you'll find the full of lots of different things. Some would go so far as to even include those things I mentioned that aren't normally considered part of A.I. However, in general, I would say that A.I. is the field that is concerned with
1) Solving the search problem (searching for a solution in a large set of possibilities)
2) Doing it with heuristics.

I'd like to take a moment to note that a famous computer vision paper came out in the 80's that documented a method called Marr-Hildreth, which was for finding edges in images. They created it by using the same technique that eyes use (laplacian of a Gaussian for edge detection - they studied cats to find this out).

A few years later someone improved upon it by throwing out the model completely and NOT doing it the way that people do (Canny).

Cognitive scientists are usually more concerned with getting the machines to do what we want than they are with modeling human thinking techniques.

Re:Not all cognitive scientists do that.

[willis]

Cognitive scientists are usually more concerned with getting the machines to do what we want than they are with modeling human thinking techniques.

I think the answer is somewhere in the middle. My experience with cogsci is that it's really about understanding thought, not about making machines. I think it really depends on where you are. If you're at MIT, it's probably machines. If you're at Berkeley, it's probably thought (at least for me, I took a class on cognitive metaphor, and we had lots in that direction. I think Santa Barbara is also more brain-focused (so-called "west-coast school")).

Re:Not all cognitive scientists do that.

[StrawberryFrog]

The goal of all the cognitive scientists I've met is to make machines think, just as with A.I.

I have never met any cognitive scientists, but I've read books on the subject by Danniel Dennet (who is arguably a philosopher not a scientist) and Steven Pinker (a cognitive scientist). The works of both of them are highly recommended.

Anyway, niether of them are focused on making machines think, but rather on understanding what makes humans think.

Re:Not all cognitive scientists do that.

[r]

trying to define AI is always problematic. very much like trying to define philosophy. :)

the classic sense of AI might have been that of search and planning. but for the last 20 years or so, many non-search and non-symbolic approaches have been treated as equals in the discipline, including:

• behavior-based robotics
  
• affective computing
  
• software agents
  
• ...and of course particular techniques like neural networks, bayes nets, markov model approximations, etc.
  

castelfranchi's introductory article in that issue actually mentions the various schisms against classic AI, which have come to be successfully reconciled with and included in the discipline.

but your're absolutely right, cog sci is more concerned with mimicking human cognitive processes. which is why AI cannot simply be a branch of it. :)

Re: Not all cognitive scientists do that.

[Black Parrot]

> The goal of all the cognitive scientists I've met is to make machines think, just as with A.I.

You need to meet more then. Ask linguists whether they're studying cog sci and they'll give you an emphatic "yes". I think these days most research psychologists would say so as well (though maybe clinical psychologists wouldn't).

> In fact, I've always heard, and was told in my AI class, that A.I. is a branch of cognitive science.

Some AI is, but not all. It really depends on the individual researcher's goals.

> However, there are many approaches to machine thinking that are not considered part of A.I.:
neural networks, SVMs, computer vision (signal interpretation), modeling.

Never heard of SVMs, but most AI researchers do think neural networks, computer vision, and certain kinds of modelling are subfields of AI.

Who taught your AI class?

> Cognitive scientists are usually more concerned with getting the machines to do what we want than they are with modeling human thinking techniques.

No, you have that backwards. AI researchers are concerned with getting machines to behave intelligently, and cog sci researchers are trying to understand human or animal cognition. And there is a fair amount of overlap, e.g. an AI/CogSci researcher may try to get a machine to behave intelligently as a model of human cognition.

Maybe slashdot could use a cognitive system...

[1337_h4x0r]

to detect dupes!

Re:Maybe slashdot could use a cognitive system...

[kwench]

Both things - detecting dupes and creating dupes - should be a simple thing today. Concerning the first one, you could easely use a spam filter, modify it, and run it over all new posts. Whenever something bears high similiarity to a former /. article, the system should print out a dupe-warning. Of course, sometimes there is wanted similiarity between different posts (like Mozilla 1.0 is out, Mozilla 1.1 is out, Mozilla 1.2 is out, Mozilla, 1.3 is out... and, guess what? Mozilla 1.4 alpha is out!)

On Combining Sensory and Symbolic Information

[slinted]

Having a system combine both symbolic logic systems and sensory systems is mentioned in the article as a major focus of research today, but I wonder why this has been split so specifically...maybe someone can help me to understand.

The point at which an understanding of body position is integrated with an overall structure of behavior leading towards a goal seems a mirage, since this isn't necessarily the way animal systems work. The best recreation of natures flexibility in "simple" systems that I've heard of comes from Mark Tilden's analog systems that are controled by tight-loops of feedback that very closely model reflex circuits, but that are capable of recovering from intense deformations of "perfect positioning".

Now, obivously, reflex systems can only go so far, when you have a bot that you want to decide path across a room, there has to be a symbolic understanding of its environment. But it seems to me, from my (albeit very limited) understanding of insect / lower-animal inteligence, that most insects don't actually work up a full symbolic understanding of their surroundings, they just have some sort of sense of direction towards a goal (think moths to light) and then they start the reflex circuits firing to move towards it. I can understand having an end goal of having a full cognitive system comparable to human understanding of the world, but it seems like people might be overshooting the process a bit. We need a greater understanding of the simple systems before we can hope to frog-leap to the big stuff.

To dispute my own point though, I feel its fair to say that the "simple" systems of the animal brain are already currently being modeled to the point that prosthesis for the brain might just be within reach. The success of an artificial hipocampus will prove that modeling the brain isn't necessarily understanding the brain, but it might be easier to learn the systems from our artificial models than the real ones.

Comment removed

[account_deleted]

Comment removed based on user account deletion

Inspired

[maharg]

The ultimate goal of the RoboCup project is by 2050, develop a team of fully autonomous humanoid robots that can win against the human world champion team in soccer

Now THAT's a goal.

Maybe we'll see humanoid robot referees in sports. That should stop any dissent from the players ,-}

Player: C'mon ref, that was never in a million years a f**king penalty !!
Ref: You have 3 seconds to comply..

can't resist

[StuartFreeman]

Johnny five is alive!

Somewhat Relevant Plug...

[Yoda2]

Experience-Based Language Acquisition (EBLA) is an open source software system written in Java that enables a computer to learn simple language from scratch based on visual perception. It is the first "grounded" language system capable of learning both nouns and verbs. Moreover, once EBLA has established a vocabulary, it can perform basic scene analysis to generate descriptions of novel videos.

A more detailed summary is available here and this is the project web site.

Compared to proprietary systems such as Ai's HAL, Meaningful Machines Knowledge Engine, and Lobal Technologies LAD, EBLA is the only system to incorporate grounded/perceptual understanding of language.

What I'd like to hear more about

[truthsearch]

While this is all very interesting and becoming more practical for everyday use, we don't hear enough about the stuff that's related but not quite bleeding edge. We know there are people trying to create intelligent systems such as for language understanding and intelligent web searching, but it seems we don't hear much about them. I'm wondering if it's because most of that is being done within corporations while much of this bleeding edge research is done by universities.

Re:Cognitive Science

[barryfandango]

Reading this reminds me of my cognitive neuroscience/AI prof Lev Goldfarb. He began our course by telling us that very, very little has been accomplished in the fields of Cog Sci and AI, and that he is possibly the only one who has brought a real contribution to the table: a formal language ("real science") for working in this field. His "Evolving Transformation System" or ETS provides methods for measuring symbols and the differences between them, and lays the groundwork for modelling cognitive processes.

Compare this to any of the fake sciences, which can easily be itentified because they have the word "science" in them. Social Science, Cognitive Science, and so on, which talk about phenomenon but fail to create formalisms to describe them (like physics does for physical phenomena, for example.)

He's eccentric, but is he right? I don't know. You can read a summary of his work here. I never dived into this field enough to learn whether he was a revolutionary or just a big talker. I'd be interested to hear what other slashdotters have to say.

Re:Cognitive Science

[percepto]

Where neurology studies the actual chemical reactions and neural activity, cognitive science studies how the "hardware" works to achieve our thought processes.

You *almost* got it. Cog Sci approaches the mind as an information processing device and seeks to understand the algorithms (mental representations and processes) operating on the incoming data. Thus, Cog Sci is the study of the mind as software not "hardware".

This is why babies can't see, even though the optics work.

Actually, newborn babies can do more sophisticated visual processing than you might think. In the first day of life, they have a preference for looking at faces over other stimuli. Plus, if you put two TV screens up with people talking on both and a speaker in the middle that's playing a soundtrack of one of the people but not the other, babies prefer to look at the TV screen that matches the sound. Thus, babies are wired to perform some fairly sophisticated cross-modal perceptual processing from the beginning.

Not to say that babies can see THAT well-- the mylenation of neurons (kinda like insulation on an electrical wire) in the brain isn't finished until years after birth, which limits the conductivity of neural signals and therefore the babies' perceptual and motor repertoire.

The perceptual system comes pre-wired for some basic things, and then self-organizes the rest based on the statistics of visual input from natural scenes. For instance, they've raised kittens in environments with nothing but vertical stripes, and after a while, they lose the ability to perceive horizontal stripes. (Sick experiments, but informative.)

Here, kitty kitty...

----------

Hey, buddy-- Can you spare a sig?

Not a great read

[Illserve]

I was disappointed by the 5 articles I read and stopped reading. It basically reads like a catalog of the projects and techno-terms that are being performed with very little actual content.

Basically each one boiled down to: our lab does the XYZAB project and we're studying this system.

Slashdot *does* use a cognitive system....

[alienmole]

It's called CmdrTaco. It's quite a sophisticated model in some respects, but it contains all sorts of spelling bugs and a strange fixation on anime and pr0n, sometimes at the same time (although only when MrsKathleenTaco isn't watching). Perhaps due in part to these irrelevant fixations, CmdrTaco's dupe-detection capability is notoriously flaky. This new cognitive research may finally allow CmdrTaco's pattern recognition systems to be upgraded, though.

But don't get your hopes up - when they attempted to upgrade JonKatz with an expanded repertoire of once-wired-now-tired cliches, the result was disastrous, and the unit had to be retired. Some upgrades are simply beyond our current technology...

Lisp: The Next Generation!

[fm6]

Actually,the thing that annoys me most is that people associate Lisp with 80s AI,

What annoys me is that people refer to Lisp as a "fifth generation language", even though it's the second oldest high-level language (after Fortran). But that's not as annoying as calling Visual Basic a "fourth generation language" because of its database features.

All of which is a secondary result of another case of 80s hype. Declarative languages, such as SQL, were sold as "fourth generation" because they were supposed to make procedural languages ("third generation languages") obsolete. Which didn't happen of course. Declarative programming ended up supplementing older languages, not replacing them.

After a while the original meaning was forgotten. So now people call languages "4GLs" etc. to emphasize some vague claim that they're more advanced. Or because of a vague notion that 4GL has something to do with database programming. These are terms we should just stop using.