These task performances tend to be bounded in 0% and 100%, right? What happens if you logit transform before predicting? I would assume that would improve prediction by a very large amount.
I tried converting the Gopher figures to logits, still got effectively zero correlation. I can’t figure out how to embed an image here, but here’s the link:
These task performances tend to be bounded in 0% and 100%, right? What happens if you logit transform before predicting? I would assume that would improve prediction by a very large amount.
I tried converting the Gopher figures to logits, still got effectively zero correlation. I can’t figure out how to embed an image here, but here’s the link:
https://imgur.com/3tg397q