Google Reveals a Machine Learning Secret or Two

July 27, 2018

I read “Google AI Chief Jeff Dean’s ML System Architecture Blueprint.” Dr. Dean is a Google wizard from the good old days at the online ad search outfit. The write up is important because it reminds me that making software smart is a bit of a challenge. Amazon is trying to explain why its facial recognition pegged some elected officials as potential bad actors. IBM Watson is trying to reverse course and get its cancer treatment recommendations to save lives, not make them more interesting. Dozens upon dozens of companies are stating that each has artificial intelligence, machine learning, smart software, and other types of knowledge magic revved and ready to deploy.

The key part of the write up in my opinion boils down to this list of six “concerns”:

  • Training
  • Batch Size
  • Sparsity and Embeddings
  • Quantization and Distillation
  • Networks with Soft Memory
  • Learning to Learn (L2L)

The list identifies some hurdles. But underpinning these concerns is one significant “separate the men from the boys” issue; to wit:

Cost

What’s this suggest? Three things from my vantage point in rural Kentucky:

First, Google is spending big money on smart software, and others should get with the program and use its technology. The object of course is to generate lock in and produce revenue for the Google.

Second, make Google’s method “the method.” Innovation using Google’s approach is better, faster, and cheaper.

Third, Google is the leader in machine learning and smart software. Keep in mind, however, that these technologies may not be available to law enforcement, to governments which wish to use the approach for warfighting, or certain competitors.

Worth reading this Google paper. One downside: The diagrams are somewhat difficult to figure out. But that may not matter. Google has you covered.

Stephen E Arnold, July 27, 2018

Comments

Comments are closed.

  • Archives

  • Recent Posts

  • Meta