26 October 1994, 12:00, WeH 1327 "Sex, Lies, and Gradient Descent (In a Multitask Backprop Net)" A Rich Caruana Talk Rated PC -- Preliminary Chat In this talk I'll present recent results on trying to understand the mechanisms underlying multitask learning in backpropagation nets. I'll present data that show: 1) tasks share more of the hidden layer if they are more related 2) multitask backprop can "average" noise in related tasks 3) mtl bp helps discrminate useful from irrelevant features 4) mtl works when there are 120 tasks! 5) helpful related tasks do not have to be "subtasks" 6) the multitask bias results from gradient aggregation, not MDL 7) the mtl bias is not due to any factor other than task relatedness