Artificial Intuition researcher Monica Anderson will take a closer look at the scientific process and what the latest developments are in that field.
Join us for this interesting evening on Thursday 21st of January 2010 6pm at SAP Labs in Palo Alto. Please RSVP. (more info further down)
I caught up with Monica and as a little teaser for our Future Salon asked her 3 questions and here are her answers (more detailed than I thought):
1) Do you have a favorite example where the reductionist method of science doesn't work, or even leads to wrong answers?
My favorite example is understanding language. No system based on grammars or other reductionist methods will consistently perform at human level on any task requiring the decoding of the meaning of human languages since language is holistic and meaning is emergent. Algorithms like those used for web search are currently mostly counting unusual words in documents and in so doing, ignoring common but important words like "not". This leads to false positives for words that have many meanings and many other kinds of search quality problems. A decade from now we will likely view our current web search algorithms as gross reductionist hacks but right now they are the best we can do. Besides web search, machine translation and voice recognition are other examples where the poor performance directly follows from their inability to understand language. Neither really works, after decades of hard work, and they will never reach human level performance as long as we try to do it using Reductionist methods. Google's latest translation systems use non-parametric models and have been outperforming all other algorithms in major competitions.
2) Intuition some may say is something that separates us from the animals. You are working on Artificial Intuition. What is it and when do you know you have succeeded in creation one?
I actually believe the opposite: All animals, including humans, have similar, basic Intuition. We use this intuition every step we take; my theory is that intuition based skills get better with practice. If practice makes perfect, then you are using intuition. We all had to practice to learn to walk.
Better and larger brains evolved over time so the intuitive competence varies from species to species. On top of the more basic skills like navigating the world and moving around in it humans have more advanced Intuition based skills that manifest themselves in several ways:
- Humans are more effective learners than other animals.
- Humans have much better language skills than other animals
- Humans have much better reasoning skills than other animals.
"Humans are better at aping than apes". We can learn complex procedures after seeing them once, which means our mechanisms for gathering experience are more effective. Part of this is that as we mature, we learn better ways to learn. But none of these skills require reasoning; they are the result of more effective versions of Intuition (the algorithm) and more data (experience gathered over a lifetime).
At Syntience we are concentrating on the task of understanding language. Our focus is Artificial Intuition - a machine learning algorithm that attempts to determine, learn, and re-use nested patterns in streams of bytes, such as text. This includes determination of saliency and abstractions, concepts, relations, etc. But before we can get to the higher levels of language understanding we have to get the lower levels right since everything builds on all levels below. 20th century AI never got to the lower parts - the kind of understanding we share with other animals - and therefore had nothing to reason about; they were building castles in the air.
We know we will have succeeded when we can outperform the current Reductionist methods on industry standard reading comprehension tests, such as word segmentation: We remove all the spaces from a paragraph of text that the system has never seen but in a language it supposedly understands. If it can put the spaces back in 100.00% correctly, then we win. If you think this is too easy a task, challenge me at the talk. And if we can get one such test consistently right, then we can likely use the same system on *any* language understanding task.
Incidentally, improved Chinese word segmentation would immediately be a product that would create significant revenue.
3) Tell us about a holistic model at work. What problem got solved that wasn't possible via old school science?
All models are the result of some reduction of the problem. Therefore there are no holistic models; if there were, then they would be "perfect simulations", not models. Also, all Model Free methods are Holistic, and vice versa. So if the question had been "Tell us about a Model Free Method at work. What problem got solved that wasn't possible via old school science?" I could have answered:
Around 1995 a friend turned me on to the Constrained Set Coverage Problem and I wrote a short program in Macintosh Common Lisp that generated, in 20 minutes on a Mac Quadra, the same (and complete) set of answers that had taken 3 days to compute on a Cray. No analytical solutions are possible. The Cray effort used a weak Model Free method (complete enumeration) and I used a more powerful Model Free Method (a Genetic Algorithm). .
The Constrained Set Coverage Problem has been shown to be NP-Hard. So Holistic Methods can sometimes be used even if the problems are NP-Hard. Of course, there is no guarantee of success and no telling how long it takes. I succeeded because of a limited dimensionality of the stated problem. But Model Free Methods will be able to try and sometimes succeed in cases where Reductionist methods would say "it can't be done".
The original problem was discussed in Science News Magazine.
This Future Salon is going to be really interesting.
Future Salons have the following structure: 6-7pm is networking with light refreshments proudly sponsored by SAP; 7-9+pm is the presentation followed by questions and discussion.
SAP Labs North America, Building D, COIL (Co-Innovation Lab). SAP is located at 3410 Hillview Avenue, Palo Alto, CA 94304[map]. Free and open to the public. Please spread the word and invite others, but be sure to RSVP http://bit.ly/52OgZu so we know how many people to expect.
We will do our best to webcast the event again via Ustream.tv on the Future Salon Channel: http://www.ustream.tv/channel/Future-Salon
Follow us on Twitter: http://twitter.com/futuresalon Follow all twittering Future Salon speakers: http://twitter.com/futuresalon/futuresalonspeakers.