paired counterfactual image editing
A causal experimental method that generates paired images differing only in a single target factor and compares differences in a model's internal features to test whether those features encode the corresponding information.
Timeline (last 90 days)
The researchers adopted paired counterfactual image editing, making targeted modifications to the same scene (such as changing the traffic light state or removing a stop sign), and found that specific SAE features do have a visual basis and change regularly with scene changes.
Paired counterfactual image editing is an experimental method for testing whether a model's internal representations have a causal relationship with a specific visual factor, by generating paired images of the same scene that differ only in one target factor and comparing the differences in the model's internal features between the paired samples.
All Facts (2)
The researchers adopted paired counterfactual image editing, making targeted modifications to the same scene (such as changing the traffic light state or removing a stop sign), and found that specific SAE features do have a visual basis and change regularly with scene changes.
50%UnverifiedPaired counterfactual image editing is an experimental method for testing whether a model's internal representations have a causal relationship with a specific visual factor, by generating paired images of the same scene that differ only in one target factor and comparing the differences in the model's internal features between the paired samples.
50%