I always welcome ambitious goals - the fact that a problem has not been solved for half a century is absolutely no reason to call it quits. However, in the interest of spending money wisely, it's worth it to at least sit back for a moment and think about what went wrong before, and what we can do to steer ourselves towards the right path.
The article mentions revisiting fundamental assumptions, but doesn't mention a single specific thing that the team will do differently. I've studied AI on my own for some time, and the biggest problem (as far as I can tell) is that researchers can't agree on what "intelligence" means. AI research for the past fifty years hasn't been about building an artificial intelligence machine, it's been about precisely defining what "intelligence" means. Every time there was a breakthrough, after a bit of hype people realized that the program is actually pretty dumb, is a testament to the intelligence of the programmer, not the machine, and that the bar for "intelligence" simply shifts a bit higher up.
So, what exactly is this team doing differently? How do they define "intelligence", and what do they intend to build?
It seems (to me) a good starting point for intelligence would be "the average person."
Refining this definition would involve researching all relevant objective tests that can be measured accurately, then forming some sort of matrix of comparison.
Initially I would pursue the testing on a single interface - using a black box approach with a text interface, then later extending it to other fields (such as movement / navigation, visual, aural etc)
If the AI can function on par with the average control, then it could be said that it is averagely intelligent based on the test matrix.
NOTE: This is just a top-of-the-head idea, I know it is a lot more complex than I make out (how do you define learning?) but it seems a logical starting point to me.
Use current tests and results - just be careful not to feed the AI the original data.
I define information as the ability to make information, which means selecting from a low probability portion of a distribution. One precise description of this is:
1. You have a problem domain P to which the No Free Lunch Theorem applies, at least approximately. Thus, there are hard bounds on how well one algorithm can do compared to any other according to some metric.
2. Information is produced when an agent can perform significantly better than is algorithmically possible. One such metric is the compressibility of its search history.
Why is this my definition? Well, it is linked to our intuitive notion of learning and intelligence. As informally described by Hofstadter, it is an inherent ability to "step outside of the system." This means, at some time t I am behaving according to some rule set r, but at time t+c I understand the rule set and can reason about r instead of just being subject to r.
One specific result of being able to reason about a rule set is that I can take some well formed sentence, realize it can't be generated by the rule set, and use a simpler rule set to generate it. When framed in terms of Kolmogrov complexity, I'm exhibiting a general compression capability, which implies a general (though not total) capability to solve the halting problem.
Since a problem domain with structure can be compressed, this relates to my first example in that if an agent has a generally much more compressible history than mathematically expected in a (almost) No Free Lunch domain, it is exhibiting the ability to step outside of its environment's rules, reason about them, and thus compress them.
So, you can see that my definition of intelligence as the ability to create information specifies an unambiguous and measurable capability, which also happens to specify something algorithms are mathematically incapable of doing. Thus, I've have both defined intelligence and disproven the logical possibility of such an AI in one fell swoop.
BTW, this is not an original thought of mine. It is a direct result of intelligent design theory, the progeny of the absolutely brilliant William Dembski.
For one, to do things differently, I'd start without trying to define 'intelligence'. Nobody can define what a game is and still we learn, understand and say plenty of interesting things about games. Nobody can you define 'water' in a way that captures all the mental images people have when they hear that word and still we can say lots of relevant things about water. Trying to capture something like 'intelligence' with words is a foolish endeavour.
I agree with you in general (you can't get wet from the word "water"). In my experience, though, if you're building something fuzzy, it usually turns out to be a complete waste of money with no results. Human beings seem to need reasonable narrow scope to produce something useful. So while the team doesn't need to waste time on formally defining "intelligence" in a way that completely captures all its aspects, they do need to precisely define what they're building, to some reasonable degree.
Without a scope definition they can't budget their funds, their time, and their human resources. Projects like these usually result in a waste of money with nothing to show for it. Of course if they try to define what they're building, it will be intimately linked to the definition of "intelligence" (assuming they claim they're building an intelligent machine). Then someone will come along and propose a counterexample that demonstrates how the machine likely isn't intelligent at all, and cannot perform well on some problem where humans do spectacularly, thereby shifting the team's scope and definition. And so, they'll be back to square one.
How dictionaries circularly define words using other words, and how humans might learn by progressively expanding analogies starting with simple 'axioms' of body sensations like up vs. down (more on other pages):
http://members.cox.net/deleyd/politics/cogsci5.htm
Wow, these three comments have just mirrored my own recent thoughts to the greatest extent I can remember. I present to you the Infinite Curiosity Loop:
If you'd like to explore this subject in a structured way, then there's a large library of philosophical writings on the subject :). I think Hilary Putnam's essay on "Brains in a vat"[1] may be a nice starting point, from which can explore both earlier and later work.
The letter 'g' is a completely arbitrary shape. Similarly an entire word 'everything' is just an arbitrary shape (though subdivided into constituent organized blobs of arbitrary shape). Letters and words seem to be symbols/code that stand for something else, and that something else might be raw sensory neuron patterns or something (imaging an apple with your eyes shut might be a trick to trigger that raw data without needing external stimulus to do so, such as seeing a real apple). Even the "sesame street" concept of up/down might require some sort of raw inner ear balance and sight sense to experience.
Rethinking your fundamental assumptions is always a good idea, especially if progress seems to be stagnating.
Some of the assumptions might be:
- That thought occurs inside the head. To what extent is thought a process distributed across multiple individuals?
- That you can make a sharp delineation between reasoning and perception. Many AI systems assume that tasks like object recognition can be completely separated from the rest of the system as it's own module. Neuroscience, on the other hand, suggests that perception, memory and reasoning are all tightly integrated together.
- That the brain can be modeled as an electrochemical system. The mainstream view is that quantum effects play no significant role, and that Penrose & Hameroff are wrong.
IMO, one of the fundamental problems with AI research is the idea that thought is important. When a chess program keeps track of the best moves nobody really cares how that thought is stored or retrieved. What's interesting is how to measure what thoughts are useful (for a given problem) and how to generate useful thoughts efficiently (better than random.) But, when people focus on thought implying something that people do, but earthworms don't do they get into a tailspin.
Language as a portable method of conveying thought is a great subject of research. But, attempting to attack a human language in all it's glory is a huge pitfall. Nouns, verbs, additives, adverbs, intonation, tense, and a 29 years of knowledge and I frequently have no idea what someone is saying. But, linking the word ball, with the concept ball, with the sensory perception of ball might be possible today. Link that to some simple nouns like roll, toss, catch, etc and we can start a meaningful integration of language and actions.
"Part of this difficulty comes from the very nature of the human mind, evolved over billions of years as a complex mix of different functions and systems. “The pieces are very disparate; they’re not necessarily built in a compatible way,” Gershenfeld says. “There’s a similar pattern in AI research. There are lots of pieces that work well to solve some particular problem, and people have tried to fit everything into one of these.” Instead, he says, what’s needed are ways to “make systems made up of lots of pieces” that work together like the different elements of the mind. “Instead of searching for silver bullets, we’re looking at a range of models, trying to integrate them and aggregate them,” he says."
It makes sense to create a single intelligent entity by combining different types of technology that works well on specific topic. We have many good techniques that do things better than human being, why try to mimic ourselves in general (the neuron way) rather than utilize a better solution to each problem?
The problem is that all of our successes combined in AI (even including neural networks) don't add up to something even remotely resembling general intelligence. It's not simply a matter of connecting the output of a DSP to the input of an edge detector, etc.
The group's approach to build cognitive assistant might be on the right way. Rather than build standalone AI, the assistive approach encourages interfacing human brain. During the collaboration between human and the assistant, both will get better understanding about the other. We human don't need computer to work with emotion, but work by understanding our need and satisfy our need just in time.
As someone who is early in the process of considering Ph.D. programs, I'd be interested if you could expound on that. Also, any schools seem like the opposite of zombie?
(This question also goes out to anyone else that has something to add.)
There are various explanations as to the cause of death: the end of the Cold War; the humbling of the mega-monopolies which funded "blue sky" research (mainly AT&T); a general loss of faith resulting from a decades-long lack of progress. Take your pick.
In fact, the entire field of computer science has been stagnant for a while, shiny gadgets to please people with 5-minute attention spans notwithstanding:
While the CS academy has many issues, the field itself is alive and well. Google and MS both do systems research, as do several financial firms and software companies servicing the financial industry.
As a person moving from physics to CS, I personally find computing to be a very interesting place right now.
Why is my desktop computer running software crippled by the conceptual limitations of 1970s hardware? Why are there "files" on my disk? Where is my single, orthogonally-persistent address space? Why is my data locked up in "applications"? Why must I write programs in ASCII text files, and plod through core dumps and stack traces? Why can't I repair and resume a crashed program?
> Where, then, is the desktop operating system not built of recycled crud?
You're willing to look past our massive advances in optimization, control systems, search technology, computer vision, etc etc, and pretend they don't exist...
Computer Vision, for one, is making massive advances.
I just spent three weeks (class project) implementing a new algorithm to find the minimum cut of a directed planar graph in O(nlgn) time. The algorithm is actually quite elegant:
This came out of a Ph.D. thesis written in 2008, and was applied to some computer vision problems in the paper I linked above. This isn't a minor speedup or optimization... it yields asymptotically faster results.
My vision professor is fairly young, and recently did his own Ph.D. work on Shape From Shading. This is the problem of recovering 3D shape from a single image (no stereo or video). His solution used Loopy Belief Propagation and some clever probability priors to achieve solutions that were orders of magnitude better than previous work. In fact, his solution is so good that rendering the resulting 3D estimate is identical (to the naked eye) to the original (although the actual underlying shape varies, since there are multiple shapes that can all appear the same given the lightning conditions and viewing angles).
There is also a ton of interesting progress in the last two decades making functional languages practical in terms of speed (and hence useful). My advisor did his Ph.D. in this area.
The entirety of CS is not evidenced by the current state of operating systems. In fact, I'd argue that OS research at this point has less to do with computation than it does with human-computer interaction, which seems like it requires more research about humans than computers.
And even systems research that is intended for desktop operating systems can be a hard sell. David Gelernter, for instance, has had a lot of interesting ideas, but getting enough regular users to adopt these systems could be a harder task than developing them in the first place.
There is nothing new about straightjacketing computing into an "everyday household objects" metaphor. As in, "the desktop," for instance. It is a very old idea which simply refuses to die.
And here is what the late Erik Naggum had to say about "user friendliness," the ancient disease which gave us MS BOB:
"the clumsiness of people who have to engage their brain at every step is unbearably painful to watch, at least to me, and that's what the novice-friendly software makes people do, because there's no elegance in them, it's just a mass of features to be learned by rote."
"The Novice has been the focus of an alarming amount of attention in the computer field. It is not just that the preferred user is unskilled, it is that the whole field in its application rewards novices and punishes experts. What you learn today will be useless a few years hence, so why bother to study and know /anything/ well?"
I tend to agree with you. I'm personally still waiting for my intelligent compiler that fixes mistakes automatically. I mean, you would think if it can tell you that your semicolon is missing that it would at least be able to fix it without interrupting you, right?... :)
Maybe things would improve if more people volunteered time toward computer science research; I can envision something like the GNU Project, but for research rather than engineering, with proper administration, goals, tasks, and resources, to help establish purpose and vision, and attract volunteers to an overarching common goal.
Or maybe even establish something like Y Combinator for CS research, a small-scale NSF if you will. Give a small group of innovative folks a few months of funding to create something new, regardless of if it has near-term business viability.
I think evolution is the best approach, given its proven power. Maybe we need replicators (like genes, only artificial/computer related) that actually codes to build hardware as the 'phenotype'. Then evolution can have something 'real' to select from instead of just a bunch of software. Only real stuff can interact with the real world obviously (opposable thumbs, eyes, ears). With DNA, the phenotype is physical like that.. molding form like hair/muscle/bone/etc.
And it seems to me that a sophisticated sense, like eyes (that can see into the real world, not just 'seeing' inside the isolated simple second life software world model or something) is needed.
I tried working on this very thing. The biggest problem I ran into is that if we program a system to do something, it will always be waiting for our feedback.
Right now I'm looking for something like a fractal. Where we get large constructs out of a small equation.
Evolution has its pitfalls. There are things that evolution cannot create because there is no chain of gradual improvements that would lead to them. A great example of that is the wheel; no living organism feature wheels of any type no matter how useful they are. The human heart would be a lot more efficient than it is if it was "implemented" as a circular pump. I think a combination of "Evolution" "Intelligent design" in AI research would be a more ideal path.
Good point. Evolution has replication at its core so perhaps is better for creating artificial life rather than specifically searching for brain designs (there'd have to be a lot of intelligent design to select brains anyway since evolution doesn't 'want' to go in any particular direction by default other than successful replication). Even though it can't forsee or jump, evolution's mass search power might outweigh these pitfalls (its the only known algorithm so far to have proven being able to build something as complex as a human brain.. though there's all sorts of bundled 'non-brain' stuff like arms, reproduction, kidneys, etc. included, which engineering could presumably ignore). This video shows 2 'wheels' of sorts in nature (but of course living things are full of mandelbroit 'roughness' and have compromises and such, instead of evolving perfectly engineered wheels):
Those are obvious rare exceptions though, and evolution never built something as fast as a jet etc. and plenty of other stuff that would need foresight. Not sure if there is some sort of somewhat simple engineering principle behind brains/consciousness that we just havn't figured out yet, but if we knew it, maybe we could engineer it like a jet instead of using evolution. A jet seems extremely simple relative to just about any of nature's locomotion creation though.
For the comment below, interesting fractal 'amplification' idea. I've briefly thought about positive feedback loops possibly building something interesting. But mostly just as some vague analogy since I don't have enough programming knowledge to experiment much.
The article mentions revisiting fundamental assumptions, but doesn't mention a single specific thing that the team will do differently. I've studied AI on my own for some time, and the biggest problem (as far as I can tell) is that researchers can't agree on what "intelligence" means. AI research for the past fifty years hasn't been about building an artificial intelligence machine, it's been about precisely defining what "intelligence" means. Every time there was a breakthrough, after a bit of hype people realized that the program is actually pretty dumb, is a testament to the intelligence of the programmer, not the machine, and that the bar for "intelligence" simply shifts a bit higher up.
So, what exactly is this team doing differently? How do they define "intelligence", and what do they intend to build?