Artificial Intelligence: The JASON Report
January 12, 2017
Compared to documents from the US Internal Revenue Service, “Perspectives on Research in Artificial Intelligence and Artificial General Intelligence Relevant to DoD, Document JSR-16-Task-003,” released in January 2017, is a literary gem. You may be able to locate the unclassified document at this link, but Beyond Search is not guaranteeing that the link will exist when you decide to pull down the 70 page report. If you are the Washington Post, looking for a source document may be too much trouble to do some “real” research. The report was prepared by The Mitre Corporation. You know the firm’s buildings on Colshire Drive. Look for the shiny windows.
Now the report.
The document does some important explanation of what artificial intelligence means. Most discussions of smart software ignore distinctions. The Jason document does not.
If you are interested in search, information access, and related disciplines, the report embraces a number of interest areas; for example, natural language processing, machine learning, and social media analysis. If you are familiar with the hive mind drone program, artificial intelligence is an important tool for warfighting, not just serving pizza ads to a mobile phone user.
Of particular interest is the discussion of “deep learning.” This term refers to the systems and methods used to make software “smart.” Recognition of objects, situational awareness, and learning are based on specific numerical recipes and procedures. Many of these have been known for many years. Fast computing makes it possible to knit together modules in order to improve the precision, recall, and reliability of data outputs. The report touches upon manifolds, a subject which is rarely included in the popular and trade banter about smart software. (For a mental image of manifolds optimization, visualize a termite nest.)
I highlight this listing of high potential topics in smart software:
- Reinforcement learning
- Graphical models (manifolds and triply periodic minimal surfaces)
- Generative models
- Probabilistic programming.
The paper acknowledges that smart software will make use of hybrid models; that is, orchestrated sequences of methods.
The discussion of hardware makes it clear that existing computer technology is a hurdle. Innovations like Google’s tensor processing unit component. In my experience, hardware remains a problem as long as going faster depends on Von Neumann architecture.
The implications for the US Department of Defense are significant. The unclassified report identifies some mountain tops and leaves the details out of the picture.
Worth reading.
Stephen E Arnold, January 12, 2017
Comments
2 Responses to “Artificial Intelligence: The JASON Report”
-
- Subscribe to Beyond Search
Feature archive
News archive
-
Did you intend to provide a link to “Perspectives on Research in Artificial Intelligence and Artificial General Intelligence Relevant to DoD, Document JSR-16-Task-003”. Can you provide a link?
No worries. Found it.