Fixed-form variational posterior approximation through stochastic linear regression
Citations
20,769 citations
10,141 citations
1,089 citations
Cites methods from "Fixed-form variational posterior ap..."
...In (Salimans and Knowles, 2013), a similar reparameterization as in this work was used in an efficient version of a stochastic variational inference algorithm for learning the natural parameters of exponential-family approximating distributions....
[...]
1,031 citations
582 citations
References
22,840 citations
10,141 citations
9,312 citations
"Fixed-form variational posterior ap..." refers background or methods in this paper
...f t>N=2 then Set g= g+ ^g t Set C = C + C^ t end if end for return ^ = C 1g Algorithm 1 is inspired by a long line of research on stochastic approximation, starting with the seminal work of Robbins and Monro (1951). Up to rst order it can be considered a relatively standard stochastic gradient descent algorithm. At each iteration, we have t = C 1g t, where we use the subscript tto indicate the values of , Ca...
[...]
...proximation is given in Appendix D. Contrary to most applications in the literature, Algorithm 1 uses a xed step size w= 1= p N rather than a declining one in updating our statistics. The analyses of Robbins and Monro (1951) and Amari (1997) show that a sequence of learning rates w t = ct 1 is asymptotically ecient in stochastic gradient descent as the number of iterations Ngoes to innity, but this conclusion rests on ...
[...]
5,561 citations
Additional excerpts
...Bottou (2010)....
[...]
4,335 citations
"Fixed-form variational posterior ap..." refers background or methods in this paper
...…distribution qη(x) of more convenient form to the intractable target distribution (Attias, 2000; Beal and Ghahramani, 2006; Jordan et al., 1999; Wainwright and Jordan, 2008). q̂ = arg min q(x) D[q|p] = arg min q(x) Eq(x) [ log q(x) p(x, y) ] , (1) where p(x, y) denotes the unnormalized…...
[...]
...A popular method of obtaining such an approximation is structured or fixedform Variational Bayes, which works by numerically minimizing the Kullback-Leibler divergence of a parameterized approximating distribution qη(x) of more convenient form to the intractable target distribution (Attias, 2000; Beal and Ghahramani, 2006; Jordan et al., 1999; Wainwright and Jordan, 2008)....
[...]