“With admirable can-do spirit, technological optimism, and a belief in inevitability, psychologists, philosophers, programmers, and engineers are sure they shall succeed [in creating human-level artificial intelligence], just as people dreamed that heavier-than-air flight would one day be achieved. … After more than 50 years of pursuing human-level artificial intelligence, we have nothing but promises and failures. The quest has become a degenerating research program …” — Peter Kassan, computer scientist, 2006.
“It is apparent to me that the possibilities of the aeroplane, which two or three years ago were thought to hold the solution to the [flying machine] problem, have been exhausted, and that we must turn elsewhere.” — Thomas Edison, inventor, 1895.
There is a popular analogy that research in artificial intelligence is to research in cognitive science as research in heavier-than-air flying machines is to research in how birds fly. That is, AI researchers want to build an intelligent machine, but it’s not especially important that it should be intelligent in a human way. Cognitive science can be a source of inspiration for AI research, but AI does not necessarily need to emulate human intelligence.
In his attack on AI research, Kassan uses the disanalogy that research in artificial intelligence is not like research in heavier-than-air flying machines, because the first has been a failure and the second has succeeded. I have several objections to Kassan’s arguments, but my main objection is that his disanalogy does not pay enough attention to the popular analogy. That is, his disanalogy is based on a confused blending of the goals of AI research with the goals of cognitive science.
Kassan’s belief that AI research “has become a degenerating research program” is based on the mistaken idea that progress in AI must be measured by how well AI emulates human intelligence (do the wings have feathers?), rather than by how well AI performs certain tasks (does the machine rise above the ground?). As an AI researcher, I read the cognitive science literature, looking for helpful clues, but I would be more excited by success with a highly non-human approach to AI than by successful duplication of human neural mechanisms in silicon. The former kind of success would tell us much more about the nature of intelligence as a general phenomenon, whereas the latter would tell us something specifically about human intelligence.
I came across Kassan’s argument while reading Ken MacLeod’s blog. Some of my points below are taken from the comments on his blog post, An AI skeptic writes.
Here are a few cases that I take to be examples of progress in AI research:
- composing music
- making paintings
- playing poker
- driving cars
- solving analogies
- translating the gist of a web page
- making patentable inventions
I think these examples show that AI research is making progress towards the goal of human-level performance on a variety of tasks. But a typical reaction is, “The poker-playing programs cannot read faces.” This is true, but it’s not really relevant to the goals of AI research. What is relevant is that the poker-playing program is playing at a level where it can challenge human professionals. As an AI researcher, I don’t really care whether the program is using human-like methods, as long as it achieves human-like performance.
Cognitive science is also making progress. In particular, I am impressed by much of the work in cognitive science on human analogical thinking. For example, there are some deep insights in Where Mathematics Comes From and The Way We Think. I admit that we’re still rather far from turning these insights into computer algorithms, but the goal of cognitive science is understanding, not computer algorithms.
Kassan seems to merge these two goals into one: we must understand how humans think and we must mirror this understanding in our AI algorithms. When you mix the two goals together, it may appear that neither AI research nor cognitive science research are making progress, as we see with Kassan:
“Without a detailed model of how synapses work on a neurochemical level, there’s no hope of modeling how the brain works.”
“After more than 50 years of pursuing human-level artificial intelligence, we have nothing but promises and failures.”
When you separate understanding human thought from implementing AI, it’s clear that we are making progress in both cognitive science and artificial intelligence.
Kassan also makes much of the gap between connectionism and computationalism:
“AI has splintered into three largely independent and mutually contradictory areas (connectionism, computationalism, and robotics), each of which has its own subdivisions and contradictions.”
We have discussed this gap before. Peter Gärdenfors’ Conceptual Spaces: The Geometry of Thought goes a long way towards closing this gap. The basic idea is that geometry is a bridge between low-level connectionism and high-level computationalism.
There is no doubt that AI researchers in the past have seriously underestimated the difficulty of achieving human-level performance on many tasks, especially daily tasks that people find easy. However, Kassan’s dismissal of AI research is just as premature as was Edison’s dismissal of the aeroplane.
Filed under: Computer Science, Philosophy of Mind | Tagged: AI, progress, research
I was under the impression that we do have detailed reports of how synapses work at a “neurochemical” level. Specifically, we have the Hodgkin-Huxley equations which allow us to model such systems. (This is a response to the quote from Kassan)
The true gap, in my eyes, is between AI and cognitive science. Those who attempt to do work in the gray area between the two fields rarely make progress - but I agree with you, those who do research in either one or the other have been making progress, and will continue to do so. I don’t think neuroscience has been attacked with the right kind of mathematics to make serious progress (to the point where we can appease Kassan); nevertheless, progress has been made.
I was under the impression that we do have detailed reports of how synapses work at a “neurochemical” level.
Lab comes one step closer to building artificial human brain
I think that the real value in AI research (successes and failures included) is to (indirectly) arrive at a better understanding of how we think / emote / reason / believe etc. i.e. AI is at the service of Cognitive Science. Building a heavier than air flying machine is of value not so much for the development of an aeroplane but for understanding the aerodynamics of how birds / bats / dinosaurs fly.
AI is at the service of Cognitive Science.
Of course, as individuals, you are free to use AI research for the insights it gives into cognitive science and I am free to use cognitive science research for the insights it gives into AI. But there is also a research community consensus. The peer review process for AI researchers reflects the community consensus that AI papers should be judged by the (actual or potential) performance of an algorithm on a task (supported by some theory or experiments with computers). Likewise, the peer review process for cognitive science researchers reflects the community consensus that cognitive science papers should be judged by the (actual or potential) empirical support for a given psychological theory about a given cognitive capacity (supported by some experiments with human or animal subjects).
This is what Kassan does not seem to understand. He wants to impose his own standards on both research communities, but he does not recognize the existing standards, and he does not present a compelling argument for abandoning existing standards in favour of his own standards.
The true gap, in my eyes, is between AI and cognitive science. Those who attempt to do work in the gray area between the two fields rarely make progress
Yes, this is a problem. The standard solution seems to be publishing two versions of work in the gray area, one version for AI researchers and another version for cognitive science researchers, each version tailored for the standards of the target community. For example, Gentner and her collaborators have published in both communities.
The true gap, in my eyes, is between AI and cognitive science.
On the other hand, separation makes a lot of sense. Imagine if the Wright brothers had felt it necessary to put feathers on their wings, or to make them flap up and down, instead of using a propeller.
Andre Vellino: Building a heavier than air flying machine is of value not so much for the development of an aeroplane but for understanding the aerodynamics of how birds / bats / dinosaurs fly.
I would rather suggest that this is exactly the other way around. Having something “heavier than air” flying was a pretty clear idea and the problem was how. Having an “artificial intelligence” is not at all that clear and even plain human intelligence isn’t a clear cut concept, thus before the “how” we have to figure out the “what”, and by this I don’t mean anything about how the brain works but about what is the “essence” of intelligence (if I can use such a dated word).
… the “essence” of intelligence …
What if there is no essence?
g, General Intelligence Factor
g, a Statistical Myth
I don’t think this would be a “show stopper” for AI research. We can go on in the current ad hoc way, tackling one interesting task after another.
Whereas we knew what flying meant before we ever saw an airplane, I only have a vague idea of what an intelligent computer might look like. Is my current computer intelligent? Clearly, it can beat me at Chess and it can compose better music than I ever could. You make a good point, Peter, to show that computers pass all sorts of Turing tests. Meanwhile, the human brain has serious “intelligence” limitations: I can’t solve certain problems (for lack of “RAM”) that a computer can solve in seconds.
It seems to me that “intelligence” is an ill-defined concept. And that’s a serious concern.
So, in the Popper’s spirit, AI is irrefutable.
So, in the Popper’s spirit, AI is irrefutable.
AI research isn’t a scientific hypothesis; rather, it is a research programme. Therefore, strictly speaking, falsifiability does not apply. With a theory or hypothesis, we can ask whether it is supported or falsified by observations. With a research programme, I think a better question is whether it is fruitful or barren.
We may not be able to precisely define intelligence, but people continue to generate statements of the form, “A computer will never be able to do X.” Computer scientists then set out to show that, actually, computers can do X. Eventually we will either run out of X’s (in which case we can announce that AI research has succeeded) or we will hit an X that we just can’t get computers to do (in which case we can announce that AI research has hit a dead end). As long as the X’s keep falling, AI research is fruitful (assuming the X’s are nontrivial).
… the “essence” of intelligence …
Of course, readers of this blog know that the essence of human intelligence is analogy-making:
Quotes: Analogy
Readings in Analogy-Making
For example, consider the analogy between AI research and research in heavier-than-air flight …
:-)
Peter: the essence of human intelligence is analogy-making
Just wondering … Analogy between what and what?
Not that I have anything against analogy: Art and Science. But it’s an old page (2002), most of the links are dead, sorry.
Just wondering … Analogy between what and what?
I would suggest, analogy between things already learned and new experiences.
In very early development, before much has been learned, many of the analogies (if you could call them that) would be between sensory inputs and very basic innate representations that have been created through natural selection.
And occasionally, two learned behaviours or representations can be improved by spotting an analogy between them, but I think it is more common that the analogy is created in parallel with the learning of the new experience. You could even say that learning the new experience (or behaviour or representation …) is creating an analogy between it and previous experiences, which is what supervised machine learning algorithms do. In the case of human learning, the ‘feature selection’ is embodied in our sensory perception systems … (they normalize, fill in missing features with mean values, etc.).
Paul: I would suggest, analogy between things already learned and new experiences.
Right, among other things, and I agree with the rest of your comment, too. But that’s not what I meant to hint at. Whenever we humans pick up on some analogy, it’s fine to think about non-formalised concepts, because we are perfectly able to handle these informal concepts even loosely and fuzzily, but if you want a computer to do that, you need some encoding of the concepts. This is the critical point!
How do you encode concepts? This is what I refer to in my blog post: “The most basic question therefore seems to be, what do we put in a concept?” There is, of course, some research already on Formal Concept Analysis, but it seems this is not enough yet to capture all the information we put in our “human informal concepts”.
How do you encode concepts?
I prefer a spatial or geometrical answer:
Conceptual Spaces: The Geometry of Thought
Geometry and Meaning
The Geometry of Information Retrieval
Evidence in favour of this approach is that the best algorithms for measuring attributional and relational similarity (synonymy and analogy) use a spatial encoding:
TOEFL Synonym Questions
SAT Analogy Questions