for overly represented concepts, like popular brands, it seems that the model “ignores” the details once it detects that the overall shapes or patterns are similar. Opening up the vision encoders to find out how these images cluster in the embedding space should provide better insights.
Yes, and this can probably be solved by methods for fairness.
I used to believe that fairness research could be ignored, that it was all rubbish, but they at least try to do something about things like unbalanced datasets etc. I'm still not sure I totally believe in it though.
If there aren't any five-legged dogs in your trainset, it's safer[0] to just remember that all dogs are four-legged than to actually recognize and count legs. After all, you might have a few images of dogs in your trainset that are misleading enough to look five-legged (e.g. because a dog is in front of another dog).
Overrepresentation is a different source of bias. That's what gives you, say, image generators that always draw "golden 1970s sci-fi robot" as C3-PO even when given additional instructions to draw something else.
Both of these problems are manifestations of the difference between training and deployment distributions. Ok, I guess you could say that four-legged dogs are "overrepresented" in the training set, but that's because four-legged dogs are also overrepresented in reality. The deployment distribution doesn't have five-legged dogs in it. What we've done is instead concoct an adversarial distribution to force a train/deploy gap where none would exist.
Releasing the vision encoder won't help because weights are opaque. Stochastic gradient descent does not yield functional internal representations[1]; it fills the bucket of parameters with one distribution and one distribution only. We could tell if, say the vision encoder produces identical embeddings for dogs regardless of leg count, or some other counterfactuals; but not much more than that.
