A comparison of sequential learning methods for incomplete data

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Probabilistic Backpropagation for Scalable Learning of Bayesian Neural Networks (2015, neutral): “Instead, the rules above match the ﬁrst and second moments of λ , which also produces good results (Minka, 2001; Cowell et al., 1996).”
Expectation-Propogation for the Generative Aspect Model (2002, methodology): “A calculation shows that the mean and variance of the Dirichlets are matched in step (b) by using the following update toγa (Cowell et al., 1996): ma = 1 Zw(γ) γ \w a”
A family of algorithms for approximate Bayesian inference (2001, methodology): “Is it optimal to minimize KL-divergence at each step, or should we use some other criterion? The modified update inspired by Cowell et al. (1996) shows that we may obtain computational speedups with other criteria. But improvements in accuracy may also be possible, if we replace the greedy KL criter”