"Psychologist Gerd Gigerenzer argues for the power of simple heuristics over more complex models when making real-world decisions. Many results in behavioral economics that appear irrational can be understood as sensible ways of coping with complexity."
econtalk.org/gerd-gigerenze
Conversation
Replying to
Tagging who has done some very cool writing about the different schools of thought between Gerd and Gary Klein with the behavior economics worlds and where each school is more applicable
1
4
Interestingly, I’m informed by my friends in AI that with model complexity, performance dips, and continues to dip until you hit a certain threshold, at which point the model starts doing better and hits a peak. After which, it starts declining again.
1
1
You're probably referring to the extended bias-variance tradeoff curve, but if so I'm not sure if performance "starts declining again", and as usual the story is much more complicated. If you're interested you can skim the paper: arxiv.org/abs/1812.11118
1
1
Yeah I should have added the caveat that it all seems very new, and the debate isn’t settled at all. But listened to the Gigerenzer podcast at my recommendation and immediately started talking about this very paper. (With the caveat that we need more work).
The dip is in error, not performance. According to the standard understanding the error on the validation data dips and then goes back up again
2
1
And yes, I'm super skeptical about the claim about the 'double descent risk curve'.
1
2
Show replies