Learning Low-order Models for Enforcing High-order Statistics
Patrick Pletscher, Pushmeet Kohli · 2012
• Standard CRF models usually trained using simple low-order losses. • In real-world often more complex higher-order losses used for evaluation. • Goal here: Train classifier directly with this higher-order loss. • Our work introduces a higher-order loss for which we can train structured SVMs exactly. Model Train a predictor of the form fw(x) = argmin y∈Y E (y, x,w). E (y, x,w) = −〈w,φ(x, y) 〉 = i∈V ψi(yi, x;w u) + (i,j)∈E ψij(yi, yj, x;w