Sunday, December 13, 2009

NIPS

Earlier this week I had a great time at the NIPS conference. I got to meet several famous AI researchers. The best part of the conference was the poster/demo sessions. My favourite posters were related to Hinton's Restricted Boltzmann Machines. I also got a chance to talk with some people from Numenta about HTM. Google and Microsoft Research both actively participated at the conference. Google had an interesting booth unveiling research they had done in using quantum computing for image search.

Disclaimer for the following paragraph: I know nothing about quantum computing and have a bad memory, so this information is not reliable. When I walked up to the booth I was very sceptical since as far as I know quantum computers are still theoretical. From what I understood they used adiabatic quantum computers to solve an optimization task. I think introducing the image search problem is misleading because any improvement in the image search is not necessarily due to the quantum computer. They should have introduced their image search algorithm and quantum optimization algorithm completely independently of each other. I would only be impressed if the quantum computer was solving an optimization problem faster than what can currently be done on a classical computer. I asked the Google guy the following question: "I remember seeing a proof that a certain type of adiabatic quantum computer could be simulated on a classical computer just as efficiently using simulated annealing. Any comments?" His response seemed to indicate that he thought that the quantum algorithm they were using could solve problems in polynomial time which could not be done on a classical computer. In addition, he gave a cool explanation of how the quantum computer did the optimization by simulating a particle in the search space and allowing the particle to travel through walls so that it doesn't get stuck in local minima. Also, the search space is slowly changed from an easy-to-solve "bowl" to the actual objective function so that the particle converges on the global minima. The algorithm supposedly needs to be run several times since it is stochastic and doesn't necessarily find the global minima every time.
Post a Comment