BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Structured Prediction Cascades - Dr. Ben Taskar (University of Pen
 nsylvania)
DTSTART:20100610T103000Z
DTEND:20100610T113000Z
UID:TALK25064@talks.cam.ac.uk
CONTACT:Simon Lacoste-Julien
DESCRIPTION:Structured prediction tasks pose a fundamental bias-computatio
 n\ntrade-off:  The need for complex models to increase predictive power\no
 n the one hand and the limited computational resources for inference\nin t
 he exponentially-sized output spaces on the other. We formulate\nand devel
 op structured prediction cascades to address this trade-off:\na sequence o
 f increasingly complex models  that progressively filter the\nspace of pos
 sible outputs. We represent an exponentially large set of\nfiltered output
 s using max marginals and propose a novel convex loss\nfor learning cascad
 es that balances filtering error with filtering efficiency.\nWe provide ge
 neralization bounds for error and efficiency losses and\nevaluate our appr
 oach on several natural language and vision problems.\nWe find that the le
 arned cascades are capable of reducing the complexity\nof inference by up 
 to several orders of magnitude\, enabling the use of\nmodels which incorpo
 rate higher order dependencies and features and\nyield significantly highe
 r accuracy.
LOCATION:Engineering Department\, CBL Room 438
END:VEVENT
END:VCALENDAR