>From http://wired.com/news/privacy/0,1848,57354,00.html, TSA will be using neural nets to harass travellers.
Neural nets, besides having "due process problems", let you infer properties --like race--- from things that you can't or won't directly ask --eg on loan applications. Its even better than the archaic "blame it on the computer" ---which doesn't cover your ass when rules are explicit. When its "blame it on the (opaque) neural net" you can get away with behavior that would not be acceptable if expressed by explicit rules. (And we have spoken to an HNC engineer who confirmed this observation.) Excerpt: The agency will spend $80 million this year to replace the current system's triggers, such as a one-way ticket purchase paid in cash, with neural networks, or fuzzy logic decision trees, that can detect more subtle indicators of a potential threat. The idea of computers acting on hunches vexes Tien. "The holy grail is that these systems will learn and adjust their suspicion calculators on their own, untethered from human input," he said. "But if you can't document the basis for a score or a decision, then you have a serious due process problem." The TSA awarded preliminary grants last spring to Lockheed Martin, Infoglide Software, Ascent Technology and HNC Software, none of which responded to requests for comment.
