AI Success Stories
I’ve been invited to give a talk on AI Success Stories, so I’ve compiled a list of things that illustrate progress in AI research. By success, for the purpose of this talk, I mean something that is interesting and impressive to a wide audience, rather than something that is successful in terms of commercial or industrial impact, or successful in terms of abstract mathematical or scientific significance, although I may mention a few examples of this other kind of success. Please let me know if there is anything you would add to this list.
Games: chess, checkers, poker, bridge, backgammon, 20 questions, scrabble, go, rock-paper-scissors.
Physical skills: driving a car, driving a motorcycle, riding a bicycle, swimming, flying a plane or helicopter, playing ping-pong, playing soccer, walking, vacuuming, cutting grass.
Art: painting, composing music, performing music, making sculpture.
Science and technology: patentable inventions, passing a chemistry test, passing a mechanical ability test.
Language: machine translation, speech recognition, character recognition, sentiment analysis, passing an analogy test, passing a synonym test.
Vision: face recognition, face detection, motion tracking, 3D from 2D.
Commerce and industry: page rank for searching, fraud detection, stock market investing.
Filed under: Computational Linguistics, Computer Science, Philosophy of Mind | Tagged: AI, progress, success
Great list. Spam filters also count, I think.
Most impressive demos I’ve seen:
flying a helicopter upside-down: http://heli.stanford.edu/
handheld speech-to-speech translation:
http://www-2.cs.cmu.edu/%7Eawb/papers/naacl2003/speechalator/
http://www.cs.cmu.edu/~awb/papers/eurospeech2003/speechalator/
Depends on what constitutes AI, but in a broad sense, Pascal’s machine was a form of AI in the sense that people did not think that machines could do accounting back then.
Hackers and researchers breaking CAPTCHAs is an interesting one (vision): problems that were designed to differentiate human beings from computers failed! (Maybe it can be described as “character recognition”.)
Grammar checking in natural languages is probably an AI problem.
Ischemia detection is one of many AI-like medical applications: a computer is able to tell that you are having a heart attack.
Theorem proving, doing algebra by computer… that’s AI, I think. I have not computed a derivative in many years.
Voice recognition is another one.
Recommender systems (such as Amazon) responsible for a large share of the sales is an automation of the “salespeople”.
Unrelated:
Will you share the slides or the notes you are preparing? As a “citeable resource” maybe?
Depends on what constitutes AI, but in a broad sense, Pascal’s machine was a form of AI in the sense that people did not think that machines could do accounting back then.
Yes, this is the well known joke that AI is what the computer cannot yet do! Accordingly there will never be any artificial intelligence, only more sophisticated algorithms. Or could some researchers provide a non-vacuous definition of artificial intelligence? I mean, not like Hutter’s: The science of Artificial Intelligence (AI) might be defined as the construction of intelligent systems and their analysis.
definition of artificial intelligence
I assume the “artificial” part is not difficult. Here are 70 definitions of “intelligence”:
http://www.vetta.org/definitions-of-intelligence/
Selmer Bringsjord has a nice discussion in his forthcoming article on AI for Stanford Encyclopedia of Philosophy:
http://kryten.mm.rpi.edu/SEP/index8.html#2
Here are 70 definitions of “intelligence”:
LOL, doesn’t this show that in fact there is none upon which one can rely? Even when there are controversial views in some scientific domain you may have 3 or 4 competing theories not 70! But the main drawback is that all those “definitions” are not stated in any scientific language but in murky layman terms reminding of virtus dormitiva.
To simplify the question and dispense with all the metaphysical questions about consciousness, symbol grounding, autonomous AI, etc, etc…. let’s assume that some day we could properly ascribe intelligence to a computer program, in which ways would this program be different from the current clumsy attempts? That is, how could we characterize the relationships between its inputs and its outputs which will allow us to label it “intelligent”?
P.S.: Selmer Bringsjord is very good at “nice discussions”, but I am interested in AI proper, not in nice discussions about AI; those interested in (not always so nice) discussions about AI should go see the Singularitarians chatter.
LOL, doesn’t this show that in fact there is none upon which one can rely?
I don’t think it is wise to “rely” on definitions. I have a very pragmatic, Wittgensteinian, Turingish, operationalist, psychometric attitude towards definitions.
If you want to do AI research, just do it. Don’t waste your time with (non-operational) definitions.
If you want to do AI research, just do it.
Sure, I can’t wait for “progress” in AI, like Deep Green for instance…
Sure, I can’t wait for “progress” in AI, like Deep Green for instance…
See Creating Friendly AI, The Second Most Important Research Problem, A Scientific Approach to Morals and Ethics, and Analogy, Ethics, Cooperation, Evolution, and the Golden Ratio.
Perhaps we could more easily define intelligence by considering the problem domains where it is required?
The AI that we can make work today is “Weak AI”. It is largely defined by the techniques used, and many books on AI are therefore catalogs of such techniques. In contrast, “Strong AI” would be able to solve a wide variety of problems in domains generally thought to require intelligence, such as true language understanding, generalized true learning and bottom-up non-instructionist world modeling. To date, there are no real success stories in Strong AI.
Programmers entering the AI field may well create excellent examples of Weak AI and may even extend the AI “bag of tricks”. A few ambitious programmers will want to work on Strong AI; but “Just do it” won’t work. Unless you understand fundamental and relevant issues in biology, philosophy and epistemology you will flail and fail. This is because most programmers will insist on using logic but logic will fail in any problem domain requiring intelligence.
Note that I am not saying these kinds of problems cannot be solved. They just cannot be solved using logic. You need intuition. In fact, I’d like to be more specific and define a Strong AI system as “Any system that can succeed on several disparate problems in problem domains that require intuition”.
Fortunately, intuition is straightforward to implement in computers.
I discuss this in more detail at http://artificial-intuition.com (5 pages).
I discuss this in more detail at http://artificial-intuition.com (5 pages).
Not really, it sounds more like marketing for your pet project, it doesn’t give the slightest idea of what you mean by “intuition”.