A sequential dual method for the structured ramp loss minimization
Dejan Mančev · SCIndeks · 2015
The paper presents a sequential dual method for the non-convex structured ramp loss minimization. The method uses the concave-convex procedure which transforms a non-convex problem iteratively into a series of convex ones. The sequential minimal optimization is used to deal with the convex optimization by sequentially traversing through the data and optimizing parameters associated with the incrementally built set of active structures inside each of the training examples. The paper includes the results on two sequence labeling problems, shallow parsing and part-of-speech tagging, and also presents the results on artificial data when the method is exposed to out layers. The comparison with a primal sub-gradient method with the structured ramp and hinge loss is also presented.