Wednesday, July 20, 2005
More on Ai and brains
It is interesting, given the topic of my last column, that the AAAI (American Association for Artificial Intelligence ) is "celebrating" the advances in artificial intelligence (AI) during the last 25 years. According to a recent PittsburghLive.com article, progress has been made. Strangely lacking in this celebration of machine intelligence is any prediction of when "real" Ai will finally appear.
This in itself is interesting. During the last 50 or so years in which people believed that computers were the "final" model for intelligence, the AI community has been fond of saying that artificial intelligence was "10 years off" 0r "just around the corner". Apparently, the community has finally learned its lesson - no prediction of when "real" artificial intelligence will exist is made by AAAI.
One can only see this as a cryptic admission that the future of computer-based intelligence is far more murky than it once was. Gone are the confident predictions of near-term success, as well as the arrogant commentary that minds are "computers made of meat". True, what some call "artificial intelligence" allows machines to do things that were thought to be the property of brains - chess-playing for example. But it is also significant that machines began playing chess 45 years ago. What have you done for us lately?
Another area that has shown some progress is "machine vision". For a long time, people who worked on machine vision were a separate breed from those working on robots. This fit with the idea that AI could work through a "narrow" sensor channel to the world. In the case of machine vision, much effort was expended on machines processing a single still image, and re-constructing the outside world from it. Outside of industrial and "toy" lab environments this approach is largely a failure. More recent work for robots centers on detecting simple primary qualities of a visual scene - movement, optic flow, 3D position. Intelligent interpretation has mostly been put on hold, with the exception of a few areas like human faces.
While there has been some success in these areas, others are more or less complete failures. After 15 years of work, the CYC project has developed an interesting "common sense" database - but hardly anything remotely intelligent. CYC programming remains custom "brain surgery" - the system can't learn on its own. Chat-bots build using the technology are supposed to pass the "Turing test" - but a few minutes with these programs in an unstructured environment shows that is false.
CYC is a big example of the field of "expert systems" - long thought to be the "golden road" to true AI. However, expert systems today are no smarter than they were in the 1980s, despite huge increases in computer power since then. Similar problems plague the field of web-based searching - despite lots of work by well-funded companies like Google, our web searches are hardly "intelligent".
Where does this leave robots? I have maintained that creating a sensor-rich body is the key to robotics - far more than building machine intelligence. It implies that we may never build robots with intelligence using computers. However, we will be able to create useful machines using current computer and AI work - provided the tasks are simple and the system has plenty of environmental feedback. In this light, a GPS-controlled robotic car seems quite possible. But a robot car that you can talk to seems a far away as ever.
It's actually liberating to think that computers can't be brains. It implies that there is another kind of machine, not a computer (but possibly comparable in the way a telephone switchboard is comparable to a Von Neuman computer) but still physical which could harbor real intelligence. Despite speed increases computers are basically the same as 50 years ago - our innovation has been around the edges, with few new "big ideas". Figuring out how the brain works may snap us out of our "computers ever faster" funk and focus on discovering some really new stuff.
It is interesting, given the topic of my last column, that the AAAI (American Association for Artificial Intelligence ) is "celebrating" the advances in artificial intelligence (AI) during the last 25 years. According to a recent PittsburghLive.com article, progress has been made. Strangely lacking in this celebration of machine intelligence is any prediction of when "real" Ai will finally appear.
This in itself is interesting. During the last 50 or so years in which people believed that computers were the "final" model for intelligence, the AI community has been fond of saying that artificial intelligence was "10 years off" 0r "just around the corner". Apparently, the community has finally learned its lesson - no prediction of when "real" artificial intelligence will exist is made by AAAI.
One can only see this as a cryptic admission that the future of computer-based intelligence is far more murky than it once was. Gone are the confident predictions of near-term success, as well as the arrogant commentary that minds are "computers made of meat". True, what some call "artificial intelligence" allows machines to do things that were thought to be the property of brains - chess-playing for example. But it is also significant that machines began playing chess 45 years ago. What have you done for us lately?
Another area that has shown some progress is "machine vision". For a long time, people who worked on machine vision were a separate breed from those working on robots. This fit with the idea that AI could work through a "narrow" sensor channel to the world. In the case of machine vision, much effort was expended on machines processing a single still image, and re-constructing the outside world from it. Outside of industrial and "toy" lab environments this approach is largely a failure. More recent work for robots centers on detecting simple primary qualities of a visual scene - movement, optic flow, 3D position. Intelligent interpretation has mostly been put on hold, with the exception of a few areas like human faces.
While there has been some success in these areas, others are more or less complete failures. After 15 years of work, the CYC project has developed an interesting "common sense" database - but hardly anything remotely intelligent. CYC programming remains custom "brain surgery" - the system can't learn on its own. Chat-bots build using the technology are supposed to pass the "Turing test" - but a few minutes with these programs in an unstructured environment shows that is false.
CYC is a big example of the field of "expert systems" - long thought to be the "golden road" to true AI. However, expert systems today are no smarter than they were in the 1980s, despite huge increases in computer power since then. Similar problems plague the field of web-based searching - despite lots of work by well-funded companies like Google, our web searches are hardly "intelligent".
Where does this leave robots? I have maintained that creating a sensor-rich body is the key to robotics - far more than building machine intelligence. It implies that we may never build robots with intelligence using computers. However, we will be able to create useful machines using current computer and AI work - provided the tasks are simple and the system has plenty of environmental feedback. In this light, a GPS-controlled robotic car seems quite possible. But a robot car that you can talk to seems a far away as ever.
It's actually liberating to think that computers can't be brains. It implies that there is another kind of machine, not a computer (but possibly comparable in the way a telephone switchboard is comparable to a Von Neuman computer) but still physical which could harbor real intelligence. Despite speed increases computers are basically the same as 50 years ago - our innovation has been around the edges, with few new "big ideas". Figuring out how the brain works may snap us out of our "computers ever faster" funk and focus on discovering some really new stuff.
