Re:Gitlab: don't discuss politics at work
by repolfx on Saturday October 19, 2019 @09:58AM
(#21298484)
I give you points for optimism and arguing in good faith :)
Unfortunately the extract I quoted is very clear. They aren't talking about understanding that engineer is a person adjective and thus could refer to an entity of unknown gender, which is a separate subfield of AI to word vectors (it'd entity analysis/knowledge graph). They're talking pure probabilities here: given a sequence of words, what is the most probable following sequence? The model gets the answer correct but Googlers are weak and cannot handle the truth, so in an Orwellian twist they label measured reality "biased" and set out to edit the model to convince it that there's no difference in probability between "Yes he did" and "Yes she did".
Already a way to debias a model (I know nothing of ML in specific, it is just an overview) would be to have the model explicitly realize that "engineer" need to be assigned a gender in certain situations. So in this sense the correct answer should be to try and understand whether this engineer is a specific human whose gender you should know so not to make wrong assumptions.
Whether this defense applies to this specific case I do not know, my point is that sometimes it is possible to actually partly debias something. The fact that it can be done in the wrong way does not mean it should not be done.
The same way there is a nice middle point between corporate anarcho-capitalism and totalitarian regimes.