IBM AI: Speeding Up One Thing, Ignoring a Slow Thing

December 12, 2017

I read “IBM Develops Preprocessing Block, Makes Machine Learning Faster Tenfold.” I read this statement and took out my trust Big Blue marketing highlight felt tip:

“To the best of our knowledge, we are first to have generic solution with a 10x speedup. Specifically, for traditional, linear machine learning models — which are widely used for data sets that are too big for neural networks to train on — we have implemented the techniques on the best reference schemes and demonstrated a minimum of a 10x speedup.” [Emphasis added to make it easy to spot certain semantically-rich verbiage.”]

I like the traditional, linear, and demonstrated lingo.

From my vantage point, this is useful, but it is one modest component of a traditional, linear machine learning “model”.

The part which suck ups subject matter experts, time, and money (lots of money) includes these steps:

  1. Collecting domain specific information, figuring out what’s important and what’s not, and figuring out how to match what a person or subsystem needs to know against this domain knowledge
  2. Collecting the information. Sure, this seems easy, but it can be a slippery fish for some domains. Tidy, traditional domains like a subset of technical information make it easier and cheaper to fiddle with word lists, synonym expansion “helpers”, and sources which are supposed to be accurate. Accuracy, of course, is a bit like mom’s apple pie.
  3. Converting the source information into a format which the content processing system can use without choking storage space with exceptions or engaging in computationally expensive conversions which have to be checked by software or humans before pushing the content to the content processing subsystem. (Some outfits fudge by limiting content types. The approach works in some eDiscovery system because the information is in more predictable formats.)

What is the time and money relationship of dealing with these three steps versus the speed up for the traditional machine learning models? In my experience the cost of the three steps identified above are often greater than the cost of the downstream processes. So a 10 percent speed up in a single process is helpful but it won’t pay for pizza for the development team.

Just my view from Harrod’s Creek, which sees things in a way which is different from IBM marketing and IBM Zurich wizards. Shoot those squirrels before eating them, you hear.

Stephen E Arnold, December  12, 2017

Comments

Comments are closed.

  • Archives

  • Recent Posts

  • Meta