|
|
|
|
|
|
by Barrin92
2242 days ago
|
|
|
The issue here is that grammatical gender is not useful information when reasoning about the semantics of some sentence. Toothpaste in German is grammatically female, that's not a reason for an ML system to make feminine assumptions about toothpaste after it combs through data, it has nothing to do with progressive values, it's that the ML system cannot distinguish between a spurious correlation and actual meaning. Today many more medical graduates are women, this will change the ratio in the future and the inference from grammatical gender will be wrong. We should be able to tell an intelligent system from the get-go to ignore something we know to be spurious rather than fiddling around with the data. And it's not strange at all to demand of an automated system that it behaves exactly the way we want it to behave. It is not human, it has to be more precise because it is rolled out at scale and it needs to do what we tell it to do. When we use industrial machinery in manufacturing we don't go "ah well humans are only precise down to a centimetre, guess we'll let it slide". An automated car trained on speeding drivers must not learn to speed. Automated systems are faster than humans, so errors compound, which requires more precision on part of a machine. If a ML system accidentally learns that ignoring someone wearing a green shirt is okay, and that is rolled out to a million cars, you don't have an accident but a big disaster. We need a way to interface with ML systems in ways that let us put precise limits on when it makes inferences from data, why it made those inferences and when to follow logical rules, and when to dispose of certain data. |
|
|