Has anyone noticed that there are a lot of families that will assign a gender identity to their child that is not the birth biological sex? I’ve encountered many on the playground with my kids. More girl-looking boys than boy-looking girls. My mind sometimes just cannot grasp why parents of perfectly beautiful girls, cute as a button and all, will insist that they’re boys, and the kids dutifully repeats the assignments, sometimes forcefully, after their parents says the sex.
(I should interject here, there is still a self in me that complains about malicious perturbations that fool our evolutionarily developed senses and instincts. The first time I mistook a stripper for a girl in a strip club was very embarrassing for me. For evolution this matter of reproduction is so important there are a lot of other auxiliary brain and body functions that serve to support sexual reproduction. I mean most sisman finds other men’s genitalia slightly to extremely repulsive. It interferes with his function. And obviously there are even more extremists in my brain that are sneering at the non-suggestion that a “boy” can be more virile than a bonafide unquoted biological boy when it comes to matters of reproductive insemination…, but let’s acknowledge their (potentially rightful) existence, but also recognize their nonsensical emotional charges chauvinistic nature… and move on)
But whatever the motivation. This sexual misdirection is actually quite good. Recall in deep learning we found it beneficial to either dropout data at many layers or else perturb the supervision like label smoothing. It improves the knowledge and skill learned.
Perhaps these perturbations, pernicious sometimes imho, will help our humanity to learn to be more true to peoples character than paying attention to superficialities like gender and skin colors.
In terms of the amount of perturbation we want, if we were able to plan it, the popular dropout rate is 50%. But clearly, at that level of perturbation, the society will cease to function in the same efficient way it does today. Label smoothing can be done with probability of 1e-3, which is 10‱ of the labels to be made not exactly right. For a school of 2000 students, that computes to 2 label altered humans. But then again, human learning is not machine learning, so perhaps we need more reinforcement.
Maybe my kids should identify as Hispanic or black… although I’d hate to be blamed for inappropriate cultural appropriations. Seems only safe thing to do is pretend to be a white American? Or opposite gender… for the sake of humanity progress.
P.s. BTW, I am not proposing that it his is a good idea normative. Labelsmoothing tend to be performed at random on separate instances. In real practice, it would mean that speaker occasionally selects a random pronoun to use instead of the actual intended pronoun. The listener will occasionally respond to a randomly selected pronoun instead of their actual pronouns. Persistently mislabel an example could cause strange biases due to population inequality. One (persistently) mislabeled black person does more than one (persistently) mislabeled white person…. Among other problems.
The FAM Blog 