That's what I was getting at; it's no longer unbiased, but it's never worse.
It's one of those cases where the obvious unbiased estimator doesn't make sense (because there's a biased one that is always at least as close to the target, and sometimes better)
@AntoniParellada That's a great figure in there, thanks!
I assumed that it's taken from somewhere, but NYT does not reference anybody and writes "We started with a set of English novels..." as if the authors of the article performed the whole analysis. Would be interesting to know the details behind how this figure was generated.
I'm new to the cross validated community and not sure how to even phrase this to look for a dupe, so I want to post this here first and if it's question-worthy, I'll post it for real. I'm attempting to use the weights of an ANN to get a rough measure of relative importance of the input vectors to the output. This is based on my understanding of Garson's approach. I'm getting a result I wasn't expecting in multiple-output neural networks:
The relative importance for all outputs is identical, but with some signs flipped.
In the last hidden layer above the outputs, the relative importance differs, but in the output layer they're all the same.
Has any one else experienced/seen anything about this? All the literature I've found so far points to single-output networks.
So I'm not sure if it's normal or if I messed up the math.
