A Smoother Way to Train Structured Prediction Models

Structured prediction is the task of forecasting highly structured outputs like sequences, where the output space is combinatorially large. A classical approach, called the structural support vector machine casts the the task of training the model as a nonsmooth minimization problem, whose first order oracle is implemented by solving a combinatorial optimization problem. In this talk, we will go over ways to smooth such a nonsmooth objective and discuss how to obtain faster nonsmooth optimization algorithms with the use of increasingly tighter smooth surrogates.