Propose a version of the subspace identity model that is robust to outliers in the training data.
Moghaddam et al. (2000) took a different probabilistic approach to face verification. They took the difference x ∆ = x 2 − x 1 and modeled the likelihoods of this vector Pr ( x ∆ | w = 0) and Pr ( x ∆ | w = 1) when the two faces match or don’t. Propose expressions for these likelihoods and discuss learning and inference in this model. Identify one possible disadvantage of this model.