The only ML algorithms with race and gender biases are those concocted by "ethical AI" researchers.
I'm talking about ML algorithms like backprop or SVMs, not the models they learn from data. (But your example is questionable anyway: if there are more whites than blacks in the population, their hairstyles will show up more.)
-
-
Tämä twiitti ei ole saatavilla.
-
What about threat bias as software error?
Keskustelun loppu
-
-
-
This is disingenuous sleight-of-hand. The lay meaning of "algorithm" is surely the behavior of a concrete trained model applied to some real problem. You're also ignoring the data-prep part of ML, which is at least as backprop and SGD and all that stuff.
-
I don't think this is disingenuous. As an ML prof, in his world, an algorithm refers to a series of computational instructions, which is of course it's technical definition. Most of these algos are designed to be dataset agnostic as well.
- Näytä vastaukset
Uusi keskustelu -
-
-
So this means that those algorithms are vulnerable to underrepresented classes and can't be used for inference where all classes are assumed equal. Although this may be "fixed" with data filtering, some kind of a "fair" algorithm may be a permanent solution to the problem.
-
For example, if an algorithm learns from balls and cubes only, it will tend to misinterpret a pyramid appearing in the data set. And for this reason the algorithm can't be used for sentencing people or something like that. This is I believe the AI Ethics team were about.
- Näytä vastaukset
Uusi keskustelu -
-
-
marketing can change wants/needs/desires
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
If the data they're meant to be returning is based on an ethical assertion (i.e., what makes hairstyles "professional" or "unprofessional"), it seems clear that they'd replicate biases. That I learnt my prejudicial bias from the data available to me does not make me *unbiased*.
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
Then the model did not learn just repeated the highest repetition of data samples
-
What? No. You are talking about memorization. If the model can account for any single data point specifically then it probably memorized the data and not learned it. Ask: given a random person, what style hair do you expect? The model will answer appropriately.
Keskustelun loppu
Uusi keskustelu -
Lataaminen näyttää kestävän hetken.
Twitter saattaa olla ruuhkautunut tai ongelma on muuten hetkellinen. Yritä uudelleen tai käy Twitterin tilasivulla saadaksesi lisätietoja.