Runaway
Wendy M. Grossman,
net.wars,
Dec 27, 2019
This is a good article despite the garish formatting (if you use Firefox you'll want to use the Just Read extension for this one). The main point of the article is that "we don't really know which patterns machine learning algorithms identify as significant." That's because the patterns aren't expressed as rules, and differences that have no meaning to us - changing a few pixels, say - may have a significant impact on the outcome. So we have to ask not only "did the model learn its task well" but also "what else did it learn?"
Today: 1 Total: 1107 [Direct link] [Share]
