Comments on: The cost of a sample http://pleasescoopme.com/2010/01/23/the-cost-of-a-sample/ Jonathan's Research Blog Fri, 20 Apr 2012 19:12:12 +0000 hourly 1 http://wordpress.com/ By: slycoder http://pleasescoopme.com/2010/01/23/the-cost-of-a-sample/#comment-137 Mon, 25 Jan 2010 20:26:40 +0000 http://pleasescoopme.com/?p=441#comment-137 Yeah, a lot of statisticians love lots of significant digits which is odd since they should be the first to recognize that many of those digits are useless.

As for SGD vs. L-BFGS, in my experience both SGD and L-BFGS are terrible for getting to lots of significant digits =).

]]> By: Bob Carpenter http://pleasescoopme.com/2010/01/23/the-cost-of-a-sample/#comment-136 Mon, 25 Jan 2010 20:16:05 +0000 http://pleasescoopme.com/?p=441#comment-136 I like the accuracy vs. time dimension. It fits in with Monsieur Gelman’s continual blog comments about sticking to justified significant digits. Especially when you look at some of his hierarchical models that have very few items from which to estimate means.

This reminds me, for some reason, of the SGD vs. quasi-Newton methods like L-BFGS for fitting regression models. SGD’s way faster at getting to a couple significant digits, but slower for getting to lots of significant digits.

]]> By: The cost of a sample « visualization, etc. http://pleasescoopme.com/2010/01/23/the-cost-of-a-sample/#comment-134 Sat, 23 Jan 2010 23:28:25 +0000 http://pleasescoopme.com/?p=441#comment-134 [...] 23, 2010 · Leave a Comment Over at “Please scoop me”, Jonathan Chang analyses the optimal number of samples one should use to estimate an expectation, [...]

]]>