Sunday, November 15, 2015

Sumlogexp, Sushi Cutting, and PTMCMC Performance

Today I Learned:
1) Numpy has a special function for calculating the log of the sum of exponentials, which is a quite numerically unstable algorithm if done naively.

2) When cutting sushi, it does indeed seem to help if you wet the blade first.

3) Parallel Tempered MCMC takes a *really* long time to run... much longer than I would have expected from a simple description of the algorithm. Specifically, it calls the objective function about an order of magnitude more frequently than I expected. =\

No comments:

Post a Comment