for “ A Practical Transfer Learning Algorithm for Face Verification ”

where the equivalency holds because p(Xi|Hi) is independent of Θt, i.e., Xi = PiHi based on Section 4 of our submission. Here p(Hi|Θt) is a zero-mean Gaussian, with block-wise diagonal covariance matrix Ωi = diag[Tμ, T , · · · , T ], where T is repeated once for each sample from subject i. Note that because the KL divergence term is independent of the hidden variables, we may ignore it for the E-step. Also, we will henceforth assume that all expectations are with respect to the appropriate posterior p(Hi|Xi,Θt) and omit further explicit reference. Using basic properties of expectations, and the Gaussian structure of p(Hi|Θt), we have