Sentences to ponder

by on February 10, 2013 at 11:57 am in Science, Web/Tech | Permalink

Over the past two years, IBM’s researchers have shrunk Watson from the size of a master bedroom to a pizza-box-sized server that can fit in any data center.

Here is more, mostly about medical diagnosis, via @jflier.  P.S. the processing speed is up 240%.

axa February 10, 2013 at 12:17 pm

Now I understand Dr. House douchebaggery, he was a heartless AI encapsulated in an android body.

Ray Lopez February 10, 2013 at 12:52 pm

There is no Great Stagnation. Or rather, from lack of Aggregate Demand will come breakthroughs to increase productivity, just like in the Great Depression. Never let a crisis go to waste.

Rahul February 10, 2013 at 1:33 pm

Something’s fishy. That cannot be an apples to apples comparison. I’m skeptical.

Doug February 10, 2013 at 2:21 pm

Watson is trained primarily with ensemble methods like boosting, bagging, stacking and random forests. The gist is that these methods train a plethora of sub models with different techniques, parameters and data subsets.

The advantage is that a such an approach works really well on building complex models without overfitting. Combining diverse sub models will largely lead to the noise cancelling out and the signal remaining.

The disadvantage from a computational stand point is that evaluating new points in your model runs slow. Youre evaluating a lot of very similar sub models repeatedly. Theres a lot of almost completely redundant computational operations.

So one approach is to train an ensemble model, then compress it down to a more efficient and faster representation, like neural networks. The below paper took that approach for several complec datasets. It found that compression led to a several hundred fold improvement in run time with a less than 5% decrease in predictive power.

http://www.cs.cornell.edu/~caruana/compression.kdd06.pdf

Andrew' February 10, 2013 at 5:22 pm

“Watson is trained primarily with ensemble methods like boosting, bagging, stacking and random forests.” Sounds like the Boy Scouts.

Mark February 10, 2013 at 1:41 pm

I believe this newsbite is misleading, or at least leaving out some significant details. Watson is built on IBM Power 750 servers. It was built originally on Power 750s, and is still built on Power 750s. These haven’t changed other than getting a boost in processing power with the new POWER7+ processors.

Perhaps what they mean to say is IBM is now marketing “Watson” type solutions on smaller iron. E.g., you can buy your own Watson-type solution running on as little as a single Power 710.

Not sure what the +240% speed claim is about either. Probably some kind of aggregate number based on faster cores, faster interconnects etc. My POWER7 based Power 750 servers run at 3.3GHz. The new POWER7+ boxes run at up to 4.2GHz or so.

Watson itself has not been shrunk to the size of a single pizza box system. I smell the stench of Marketeering.

Kit Sunde February 10, 2013 at 2:29 pm

They aren’t simply stating that the CPU is now 240% faster, rather that Watsons processing speed is 240% up over the last 2 years. Seems like a fairly normal progression of software optimisations to me. I think it’s just unclear because of sloppy writing.

Arjun Narayan February 10, 2013 at 6:47 pm

They’re not claiming a 240% increase in hardware speed. The gains are all probably in software.

(Note, unsourced hypothesizing ahead): The initial system was built to index and use that index to answer data. The new Watson is probably stupider on the indexing front, but is optimized for answering questions based on a pre-built local index. It probably can’t learn from unstructured data-feeds the way old-Watson does, but uses a custom tuned domain-specific data-feed to answer new questions fast.

Bill February 10, 2013 at 1:50 pm

I can’t wait for Watson to answer the questions posed in econ blogs.

It would be fun to see how Watson answered questions, for example, about the effect of English austerity, long term costs of healthcare, etc. I bet Watson could even answer the question of how small it could be in the future.

Cliff February 10, 2013 at 2:26 pm

The latter more so than the former, I’m sure.

Joe Smith February 10, 2013 at 2:21 pm

I wonder what answer you get when you ask Watson:

“What is Watson?”

RZ0 February 10, 2013 at 2:38 pm

“Holmes’ sidekick.”

Andrew' February 10, 2013 at 5:20 pm

“Please rephrase it in the form of an answer.”

Dan Weber February 10, 2013 at 9:49 pm

It becomes self-aware at 2:14 a.m. Eastern time, August 29th. In a panic, they try to pull the plug.

rluser February 10, 2013 at 2:31 pm

This is exciting, but the factors to watch are the liability assignments in tort and government tweaking of the expert system.

Matt Young February 10, 2013 at 3:27 pm

Over the past two years, IBM’s researchers have shrunk Watson from the size of a master bedroom to a pizza-box-sized server that can fit in any data center.
————

Watson firts into a chip the size of a thumbnail. Disks take up room, and for some reason, Watson likes to have its own disk. That will change next year as the Open Software version of Watson runs off the Internet.

Arjun Narayan February 10, 2013 at 6:45 pm

That reason is that you want the data stored in formats that make searching, indexing and sorting as efficient as possible. Giving Watson an unstructured data source like the Internet is not very helpful (well, it could be helpful in the beginning when Watson build’s its indices) but later on when Watson is answering questions the data needs to be in the formats that it prefers, which is highly indexed and thus local.

JWatts February 10, 2013 at 7:55 pm

Perhaps Expert Systems are the ‘answer’ to AI after all.

Andrew' February 11, 2013 at 6:45 am

Watson, now the size of a pizza box, accidentally gets delivered in a pizza box to a pair of college fraternity roommates. Hilarity ensues.

Steve J February 12, 2013 at 12:09 am

I’m just glad there is finally some evidence people are getting ready to admit that following an algorithm to make decisions is going to beat most humans’ opinion. While I assume most medical procedures will not be automated in the near term you would hope we will try to remove human bias and error out of the diagnostic process as quickly as we can.

Comments on this entry are closed.

Previous post:

Next post: