4 ms·

The problem with calling something biased is that it carries a connotation of racism. There is a difference between disparate impact and disparate treatment, t

by mises 7y ago

The problem with calling something biased is that it carries a connotation of racism. There is a difference between disparate impact and disparate treatment, to use Eric Holder's terminology. I don't think humans are "projecting their own experiences", I think that the dataset that is fed to algorithms over-represents whites because they constitute a larger portion of the population. It may also over-represent asians because they are a larger portion of the silicon valley labor force. This does not mean they are attempting to mis-treat blacks, hispanics, etc. Maybe there is progress to be made, but it will be difficult to make a facial-recognition algorithm work as well on the 13% of the population that is black compared to 77% white. That is, until the third world becomes a market. Then, they will be tremendously biased to non-whites, though I don't consider either bad.

kaitai 7y ago
"Biased" still has real statistical meaning. It is unfortunate that people treat these problems with an aim to justifying their virtue rather than creating a product that works (not saying you are doing this, just commenting that often the reaction is about personal virtue rather than actual technical accuracy). It is not more difficult to make a facial-recognition algorithm work on black people -- or Asian people, to pick another example in the news (iPhone & users in China). Just choose your input data appropriately. All it takes is some work and a model that wasn't trained on Caucasian faces.
- mises 7y ago
  It might actually be more difficult. There have been cases where soap dispensers didn't work on blacks because more infrared was absorbed. Not saying it's not solvable, but it's possibly harder. I guess you're right that you could make different models and use the correct one based on race. Correct me if I'm wrong, I'm no expert, but could this be solved by adding an extra layer to a neutral network?
  - kaitai 7y ago
    For facial recognition I believe it's possible to use pre-trained networks and add extra layers for specific (visual) demographics. You're correct that for the engineering of soap dispensers an actually different method might be preferable. That however boosts a previous point; if the engineers behind the soap dispenser had a number of black members, would they have designed something that didn't work for them, or discarded it in the design stage?

by mises 7y ago

kaitai 7y ago

"Biased" still has real statistical meaning. It is unfortunate that people treat these problems with an aim to justifying their virtue rather than creating a product that works (not saying you are doing this, just commenting that often the reaction is about personal virtue rather than actual technical accuracy). It is not more difficult to make a facial-recognition algorithm work on black people -- or Asian people, to pick another example in the news (iPhone & users in China). Just choose your input data appropriately. All it takes is some work and a model that wasn't trained on Caucasian faces.

mises 7y ago
It might actually be more difficult. There have been cases where soap dispensers didn't work on blacks because more infrared was absorbed. Not saying it's not solvable, but it's possibly harder. I guess you're right that you could make different models and use the correct one based on race. Correct me if I'm wrong, I'm no expert, but could this be solved by adding an extra layer to a neutral network?
- kaitai 7y ago
  For facial recognition I believe it's possible to use pre-trained networks and add extra layers for specific (visual) demographics. You're correct that for the engineering of soap dispensers an actually different method might be preferable. That however boosts a previous point; if the engineers behind the soap dispenser had a number of black members, would they have designed something that didn't work for them, or discarded it in the design stage?