3 ms·
> Yes, but those structural priors have nothing to do with race, gender, or any other protected attribute. This seems like wishful thinking. If fed a data set
by seoaeu 5y ago
> Yes, but those structural priors have nothing to do with race, gender, or any other protected attribute.
This seems like wishful thinking. If fed a data set containing 'race' or 'gender' as one of the fields, most models have structural priors that will make them assume all correlations between race/gender and other fields are meaningful. Worse, just because an input data set doesn't have race or gender recorded doesn't mean that the model won't predict them, and then use the results of those predictions to bias its output
- darawk 5y agoDon't conflate "meaningful" with "predictive". The attributes are indeed predictive. The intention of the model is to make accurate predictions optimally given the data its fed. If you give it data that contains predictive correlations with race and gender, yes, it will learn them. It should be pretty clear that the root problem there though is the data, not the model.