StrandedNomad on Nostr: Learned something interesting about the use of AI for generating counterfactual ...
Learned something interesting about the use of AI for generating counterfactual examples.
An interesting implementation was that Google’s AI model was able to use a picture of the fundus to predict with near 100% accuracy the gender of the individual whose eye it was. Ophthalmologists and optometrists cannot predict the gender on the basis of a picture of the fundus, they can only guess.
So what they did was feed pictures of the retina to a local AI model. After being fed this dataset, it was asked to take a real picture of someone’s retina and create a picture of what the retina would look like if the subject had the opposite gender. (A counterfactual picture)
What they found was that the AI generated a fovea (the deepest part of the retina) deeper in males and shallower in females. Thus, they found out that the original AI model was able to determine the depth of the fovea and then make a prediction about the gender of the subject with near 100% accuracy.
An interesting implementation was that Google’s AI model was able to use a picture of the fundus to predict with near 100% accuracy the gender of the individual whose eye it was. Ophthalmologists and optometrists cannot predict the gender on the basis of a picture of the fundus, they can only guess.
So what they did was feed pictures of the retina to a local AI model. After being fed this dataset, it was asked to take a real picture of someone’s retina and create a picture of what the retina would look like if the subject had the opposite gender. (A counterfactual picture)
What they found was that the AI generated a fovea (the deepest part of the retina) deeper in males and shallower in females. Thus, they found out that the original AI model was able to determine the depth of the fovea and then make a prediction about the gender of the subject with near 100% accuracy.