FF-LEACE (Free Form LEACE) is a new closed-form concept erasure method that extends the original LEACE approach. Unlike oracle LEACE, it does not require access to concept labels at inference time, and avoids injecting non-linearly represented information into representations. The method finds the function nearest to the identity such that no linear classifier can predict the concept better than chance. A PyTorch implementation is available on GitHub, demonstrated on a toy problem where the conditional expectation is computable in closed form.
1 Impression