AC2005: Prospects for AI panel

09/17/05 00:00:00    

By Michael Mealling

This is a panel on the Prospects for AI. Panel members are Neil Jacobstein, Patrick Lincoln, Peter Norvig, and Bruno Olshausen. First up is Neil Jacobstein talking about the current state of AI and what's real and what isn't. In general the terms have changed though: In other words, if it works, then it isn't “AI”. AI has become a real aspect of IT in a much more incremental and 'stealth' way rather than the old hyped announcements of super-brain computers that can out think humans.

Now he's reviewing the knowledge, systems, business and cultural issues still to be solved. To many to blog well, though. He's covering stuff from the Semantic Web to nanotech and reverse brain engineering. So this is really a review instead of something really new.

Now Patrick Lincoln is reviewing why you need AI and a review of the various flavors of AI in increasing difficulty (i.e. SAT, Baysian filters, Byzantine fault tolerant systems, etc). “It is more interesting to explore ALL of the behaviors of an abstract system than SOME of the behaviors of a complete system”.

“What is the worlds most urgent and important problem?” Collectively getting better at solving urgent and important problems….

Next up is Peter Norvig from Google: his first slide is “AI in the middle”. AI as a mediating force between people. His point is that humans have actually done most of the work building the knowledge base that an AI needs to infer over, its just a matter of making it accessible and searchable (ala Google). A good example is statistical machine translation that uses human translations to build an AI engine that can take over once it has learned how to translate.

Next up is Bruno Olshausen on brain modelling and applying that to real world AI. He is starting a company called Numenta to commercialize it. Title slide is “Neuroscience and Future Prospects for Intelligent Systems”. He makes the point that much of what they're learning isn't from human brains. Jumping spiders for example don't have compound eyes and have a very advanced visual system that runs on 30,000 neurons. The current state of neuroscience is still limited to the point where they still don't know what kind of computing device a neuron actually is. Theoretical neuroscience is a combination of experimental psychology, neurobiology and math/computer science.

Its clear from all of the speakers that the old “top down” model of building AI systems has lost the fight. Especially since its gotten rather obvious that building large complex systems “top down” generally doesn't work. Its certainly not the way nature has learned to do it.