Abstract: Learning theory has burgeoned since 1971 (VC-dimension by Vapnik and Chervonenkis) when the framework was established and the fundamental theorem proved. We extend the framework to model a strategic setting that allows for adversarial behavior during the test phase. As an example consider a situation where students may manipulate their application materials knowing universities’ admissions criteria. We define a new parameter, SVC (Strategic VC dimension) and use it to characterize the statistical and computational learnability of the strategic linear classification problem. The practice of learning exploded in 2012 (AlexNet by Krizhevsky, Sutskever and Hinton) leading to a profusion of deep neural network (DNN) architectures and training algorithms. We consider the problem of creating tags from a noisy scanning technology (e.g. optochemical inks, magnetic microwires etc.).
Venue: Sala de Seminario John Von Neumann, CMM, Beauchef 851, Torre Norte, Piso 7.
Speaker: Ravi Sundaram
Affiliation: Northeastern U, USA.
Coordinator: José Verschae