Alpha-divergence minimization for deep Gaussian processes

This paper proposes the minimization of α-divergences for approximate inference in context deep Gaussian processes (DGPs). The proposed method can be considered as a generalization variational (VI) and expectation propagation (EP), two previously used methods DGPs. Both VI EP are based on Kullback-Leibler divergence. is scalable version power propagation, that introduces an extra parameter α specifies targeted α-divergence to optimized. In particular, such recover solution when α→0 α→1. An exhaustive experimental evaluation shows via feasible DGPs choosing intermediate values between 0 1 give better results some problems. means one improve training Importantly, allows stochastic optimization techniques, making it able address datasets with several millions instances.