Andrew Gordon Wilson, NYU Courant professor, references his 2015 paper 'The Human Kernel' and describes an unpublished human GAN project examining mind reading through human subjects.
The paper's final section influenced an ICML 2022 best paper on model selection.
#56Dan Roy@ROYDANROYHaving not the paper, I'm curious how you distinguish the humans having a kernel and learning a kernel. The latter contains the former as a special case, but would undermine the assumptions you likely needed to extract kernel evaluations from human judgments.
It would be interesting to assume the humans are learning a kernel, and try the same exercise. Not obvious how to do it.
Andrew Gordon Wilson@andrewgwils"The human kernel" is a top contender for my most eccentric paper, a manic sprint with a crazy story behind it: https://arxiv.org/abs/1510.07389. The last paragraph of "human Occam's razor" planted the seed for our model selection work (ICML 2022 best paper) almost 10 years later. 1/2
There was a follow up, the human GAN, which was borderline crackpot, and unfortunately never saw the light of day. It turns out that mind reading and differentiating through people is hard, but possible. 2/2
Andrew Gordon Wilson@andrewgwils"The human kernel" is a top contender for my most eccentric paper, a manic sprint with a crazy story behind it: https://arxiv.org/abs/1510.07389. The last paragraph of "human Occam's razor" planted the seed for our model selection work (ICML 2022 best paper) almost 10 years later. 1/2