Re:Google promoted Texas gunman fake tweets
by ForHackernews on Thursday November 09, 2017 @06:37AM
(#15661572)
> Algorithms are generally incapable of bias.
This is nonsense unless you define-down "bias" in such a way as to mean "performing contrary to the deterministic result of its implementation" (which seems to be what you're doing, and I agree is impossible).
A more sensible definition of bias might be something like "preferentially favouring one outcome over another for reasons unrelated (or even contrary) to the core purpose of the software". Imagine a vending machine that when prompted generically by a user saying "I'm thirsty" always provided sugary soda instead of bottled water.
Modern machine learning algorithms are extremely complex and opaque, and depending on the dataset they're trained on, and the features they're trained to focus on, they can very easily wind up just encoding human biases.
This is nonsense unless you define-down "bias" in such a way as to mean "performing contrary to the deterministic result of its implementation" (which seems to be what you're doing, and I agree is impossible).
A more sensible definition of bias might be something like "preferentially favouring one outcome over another for reasons unrelated (or even contrary) to the core purpose of the software". Imagine a vending machine that when prompted generically by a user saying "I'm thirsty" always provided sugary soda instead of bottled water.
Modern machine learning algorithms are extremely complex and opaque, and depending on the dataset they're trained on, and the features they're trained to focus on, they can very easily wind up just encoding human biases.
https://www.theregreview.org/2017/10/10/cramer-bias-computer...