This preview shows page 1. Sign up to view the full content.
Unformatted text preview: earning algorithms. The user gives the algorithm many examples of desirable messages and also some counterexamples of undesired trafﬁc. “The software identiﬁes all the variables that inﬂuence the property that you are interested in [for example, not spam], then searches over all feasible relationships among those variables to ﬁnd the model that is most predictive,” Horvitz explains. Bayesian networks can be eerily accurate. “They use probabilities, so they are wise in the sense that they know that they can’t know everything,” Horvitz
JANUARY 2005 that information via a wireless link to the cell phone in his pocket, which then switches from ring mode to vibrate. Although the technology is steadily improving, gaze detectors are still too expensive, bulky, ugly and unreliable for everyday use. “Eye contact is the most accurate measure of attention that we have — about 80 percent accurate in conversational settings,” Vertegaal says. “But it’s not perfect by any means.” Attentive appliances are mere parlor tricks, moreover, when they act independently. The real payoff will only come from larger, smarter systems that can both divine the focus of our attention and moderate our conversation with all our personal machines. Doing that reliably will require a nice bit of reasoning. Trusting the Black Box
broa dly sp e a k i ng , computers can use two techniques — rules or models — to decide when and how to transmit a !
60 SCIENTIFIC A MERIC A N COPYRIGHT 2004 SCIENTIFIC AMERICAN, INC. KC ARMSTRONG “Artiﬁcial intelligence couldn’t deliver the personal secretary. I’m pretty sure we can deliver a personal receptionist.” elaborates. “That allows them to capture subtle behaviors that would require thousands of strict rules.” In January he plans to present the results of a ﬁeld trial of a model trained on 559 past appointments taken from a manager’s datebook. When challenged with 100 calendar entries it had never seen, the model correctly predicted whether the manager would attend the meeting 92 percent of the time. And in four out of every ﬁve cases, the model matched the manager’s own estimate of the cost of interruption during the meeting. That sounds impressive, but some experts in the ﬁeld remain skeptical. Users may have a very low tolerance for a system that erroneously suppresses one out of every 10 important calls. “The more ‘attentive’ things become, the more unpredictable they are,” warns Ben Shneiderman of the University of Maryland. “We have a history in this community of creating ‘smart’ devices that people don’t use because they can’t understand how they operate.” Indeed, Vertegaal reﬂects, “artiﬁcial intelligence couldn’t deliver the personal secretary, because it was too complicated.” Nevertheless, he adds, “I’m pretty sure we can deliver a receptionist.”
w w w. s c ia m . c o m That would be welcome, but will considerate computing really reduce interruptions and boost productivity? At least for certain speci...
View Full Document
This note was uploaded on 02/24/2010 for the course COMM 4400 at Cornell University (Engineering School).