Supplementary Materials for: Unsupervised Object Class Discovery via Saliency-Guided Multiple Class Learning
Jun-Yan Zhu, Jiajun Wu, Yichen Wei, Eric Chang, Zhuowen Tu · 2012
1. Proof for Theorems Now we will do discriminative learning with the presence of hidden variables. Our step is similar to standard EM[3] while the primary difference is that we are given labels Y = {y1, . . . , yn} in addition to observations X = {x1, . . . , xn}, and we want to estimate the model θ that minimizes the negative log-likelihood function L(θ;Y,X) = − log Pr(Y |X; θ). We proceed by integrating H out: