On Sat, 14 Apr 2001, Aimee Farr wrote:
That is an over-simplification, but yes. Intelligence is not headlines. To a large extent, "what's happening" is not analyzed correctly, because the intelligence community lacks sufficient expert analysis to cope with the dataload. This capability is in the private sector. These information flows, between the government sector and the private sector, are unmapped.
This is not true any more. The automated analysis of trawled data has advanced considerably beyond keyword searching at this point; there are programs out there now specifically looking for much more subtle and complicated things, which were formerly the domain of intelligence analyists, and they are actually pretty damn good. The simple keyword searchers and keyphrase searchers you hear about with echelon are only the front line; they pass their data back to much more sophisticated AI programs that analyze content, and synthesize information gleaned from massive numbers of such missives. Every time a situation like the Aum Shenrikyo (spelled?) subway attack happens, if the automated analysis suite didn't point it out first, human analysts come in and check out the dataflows that ran before it and around it, and create a new auto-analysis program. And then later, when another group that has anything like the same rhetoric and seems to be going through the same logistical steps pops up, the auto-analysis finds it without human help. I do not speak of specific known programs here; but my primary background is in AI and expert systems, and I can state unequivocally that intelligence analysis funded most of the research in the field for a very long time, and that programs such as I described above are well within the current state of the art. It is unusual for them to be deployed very widely in private industry because in private industry there is a real problem of retaining personnel with the proper expertise to work on them. They tend to be delicate in their operation -- you go to make a minor change in the data or the rules or the schemas and the performance of all other parts of the system degrades unless you are extremely careful, well-trained, and, let's face it, consistently just plain smarter than normal people. But when they are in tune, and their vocabulary tables are up-to-date, they are highly accurate. The problem of keeping these systems in tune is what drives most practical AI research today; the systems are effective, but brittle and unable to cope with subtle changes and variations very well. "Fuzzy" approaches like ANN's and Genetic Algorithms are attempts to get past this problem by making self-adjusting systems, but the volumes of data required to get self-adjustment working using such approaches are a problem; you'd have to have data from hundreds of Aum Shenrikyo type attacks before your GA or ANN really had a good chance of picking out what parts of the dataflow were relevant. So here's my speculation: human analysts are probably called in only after something takes the automatic tools by surprise, or when there is an administrative need for specific analysis that the automatic tools do not provide. Bear