Professor who says facial recognition can tell if you're gay
theguardian.com
theguardian.com
The takeaway: AIs can learn our stereotypes.
The question is then- if they believed their sources that homosexual men and women are 7% of the population, why did they choose to train their classifier with a radically different distribution. One answer may be that it's much harder to train an accurate classifier for unbalanced classes and that artificially including more of the minority class improves the performance of the classifier on its training and its test set, therefore overall improving its evaluation.
Of course, the resulting model is useless for predictions in the real world, because it violates the central assumption in PAC learning (the computational learning theory behind modern machine learning): that the data used to train a machine learning algorithm has the same distribution as unseen data.
In other words, the "deep gaydar" paper is a pointless exercise in classifier tuning. It proves that, with enough knowledge of how classifiers work, you can squeeze any kind of results you want out of them. But it doesn't tell us anything else.
That is not controversial. I'm surprised at your statement that distributional consistency is "not required at all" and I wonder where it's coming from.
I don't understand what you mean by "data around the edge of the classification set".
The images are independent, each holds a classification and how many you have from each type does not need to reflect their actual distribution.
It’s like the atomic bomb— sure it’s incredibly dangerous technology and we’d prefer no one has it. But we can’t prevent others from having it, and if everyone has it mutually assured destruction actually promotes peace.
For instance, what if the AI can read the same from children-- say even in the womb? There are already babies killed before birth because they are the wrong sex, or because they show potential to have some disease.
Caution is warranted.