Discussion about this post

User's avatar
Maxim Raginsky's avatar

"We’re most definitely optimizing, as we really care about these competitions on dataset benchmarks. But our algorithm is some sort of massively parallel genetic algorithm, not a clean, rigorous, and beautiful convex optimization method."

See? This illustrates my point about Darwin being the synthesis of Hume (messy, collective lurching about in the space of algorithms, objectives, etc.) and Kant (individual engineers finding a model on the internet and tweaking a parameter or two).

Expand full comment
Joe Sill's avatar

What's maybe ironic about that list of methods from Duda and Hart is that it although it's a long list, it doesn't appear to include a standard statistics-style logistic regression fitting procedure (i.e. maximum-likelihood a.k.a. cross entropy, I believe typically done in batch mode). So you could easily add that to the list.

Expand full comment
20 more comments...

No posts