3 ms·

I remember stumbling across a "beauty rating" site that had an ML model trained to rate photos on beauty. At the risk of sounding conceited, when it gave me (b

I remember stumbling across a "beauty rating" site that had an ML model trained to rate photos on beauty. At the risk of sounding conceited, when it gave me (black male) a bad score, I knew something had to be wrong. I went and searched for photos of famous black males generally accepted as attractive and noticed it also gave them poor scores. That gave me a hunch about what was going on, and lo and behold: taking the same exact photos, but processed to black and white and brightened, resulted in extremely high scores. The model was essentially measuring how light your skin is.

HarryHirsch 5y ago
ML models are... very local. They fail at sensemaking. If it's any comfort, the same model would probably identify President Mobutu as a leopard.
tempnow987 5y ago
There are some studies I believe that show even attractive african american women can have trouble with low click through / match / rating on dating sites. So if the model was trained on who folks wanted to date (as a measure of attractiveness) the model could still be somewhat accurate, even though not accurately measuring attractiveness. How would models work on rating who would be a good NBA player / runner etc. I wonder if just being white would get me downrated.
- BoorishBears 5y ago
  Are you missing the part where the model gave two scores for the same picture based on extreme desaturation? Terrible pictures don't make you more attractive on dating sites... It's just a bad model. There's bound to be bias in the model but ML 101 would be randomize saturation and hue on your training set
  - tempnow987 5y ago
    Fair enough. That said, I wouldn't jump to conclusions on how things like saturation or desaturation play into ratings, a lot of instagram filters for example have an a desaturation effect which makes me wonder if folks like that as a look. I don't personally. Very well could be a bad model, but sometimes there are underlying factors that result in these bad models from the training set. And no, it's no ML 101 to randomize elements of your input in the training set, though if you want a broadly applicable model you'd look to have inputs that include a variety of hue and saturation (perhaps that's what you meant - I have seen folks though manually start doing things like randomize hue in the set, where the ratings were on the original image).
renewiltord 5y ago
That shit is just some dude who put an ImageNet based model on the Internet. I remember going through one of these and getting “First time offender”. I was wearing an orange shirt. That’s a horrendous choice of colour, honestly, but I’ve never been actually caught for any of my crimes and it’s been a while since I committed them first.

HarryHirsch 5y ago

ML models are... very local. They fail at sensemaking. If it's any comfort, the same model would probably identify President Mobutu as a leopard.

tempnow987 5y ago

There are some studies I believe that show even attractive african american women can have trouble with low click through / match / rating on dating sites. So if the model was trained on who folks wanted to date (as a measure of attractiveness) the model could still be somewhat accurate, even though not accurately measuring attractiveness. How would models work on rating who would be a good NBA player / runner etc. I wonder if just being white would get me downrated.

BoorishBears 5y ago
Are you missing the part where the model gave two scores for the same picture based on extreme desaturation? Terrible pictures don't make you more attractive on dating sites... It's just a bad model. There's bound to be bias in the model but ML 101 would be randomize saturation and hue on your training set
- tempnow987 5y ago
  Fair enough. That said, I wouldn't jump to conclusions on how things like saturation or desaturation play into ratings, a lot of instagram filters for example have an a desaturation effect which makes me wonder if folks like that as a look. I don't personally. Very well could be a bad model, but sometimes there are underlying factors that result in these bad models from the training set. And no, it's no ML 101 to randomize elements of your input in the training set, though if you want a broadly applicable model you'd look to have inputs that include a variety of hue and saturation (perhaps that's what you meant - I have seen folks though manually start doing things like randomize hue in the set, where the ratings were on the original image).

renewiltord 5y ago

That shit is just some dude who put an ImageNet based model on the Internet. I remember going through one of these and getting “First time offender”. I was wearing an orange shirt. That’s a horrendous choice of colour, honestly, but I’ve never been actually caught for any of my crimes and it’s been a while since I committed them first.