sandwiching a marginal

When working recently on a paper for estimating the marginal likelihood, I was pointed out this earlier 2015 paper by Roger Grosse, Zoubin Ghahramani and Ryan Adams, which had escaped till now. The beginning of the paper discusses the shortcomings of importance sampling (when simulating from the prior) and harmonic mean (when simulating from the posterior) as solution. And of anNealed importance sampling (when simulating from a sequence, which sequence?!, of targets). The authors are ending up proposing a sequential Monte Carlo or (posterior) particle learning solution. A remark on annealed importance sampling is that there exist both a forward and a backward version for estimating the marginal likelihood, either starting from a simulation from the prior (easy) or from a simulation from the posterior (hard!). As in, e.g., Nicolas Chopin’s thesis, the intermediate steps are constructed from a subsample of the entire sample.

In this context, unbiasedness can be misleading: because partition function estimates can vary over many orders of magnitude, it’s common for an unbiased estimator to drastically underestimate Ζ with overwhelming probability, yet occasionally return extremely large estimates. (An extreme example is likelihood weighting, which is unbiased, but is extremely unlikely to give an accurate answer for a high-dimensional model.) Unless the estimator is chosen very carefully, the variance is likely to be extremely large, or even infinite.”

One novel aspect of the paper is to advocate for the simultaneous use of different methods and for producing both lower and upper bounds on the marginal p(y) and wait for them to get close enough. It is however delicate to find upper bounds, except when using the dreaded harmonic mean estimator.  (A nice trick associated with reverse annealed importance sampling is that the reverse chain can be simulated exactly from the posterior if associated with simulated data, except I am rather lost at the connection between the actual and simulated data.) In a sequential harmonic mean version, the authors also look at the dangers of using an harmonic mean but argue the potential infinite variance of the weights does not matter so much for log p(y), without displaying any variance calculation… The paper also contains a substantial experimental section that compares the different solutions evoked so far, plus others like nested sampling. Which did not work poorly in the experiment (see below) but could not be trusted to provide a lower or an upper bound. The computing time to achieve some level of agreement is however rather daunting. An interesting read definitely (and I wonder what happened to the paper in the end).

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.