LEACE
In plain wordsA one-step mathematical method for scrubbing a concept, such as gender, from an AI model's internal data so that no simple detector can find it, while changing everything else as little as possible.
Least-squares concept erasure: a closed-form method of linear concept erasureconcept erasure, n.: Removing a specific idea — such as a person's gender, or an artist's style — from what an AI model can represent or produce, so the model can no longer use or show it. that applies an affine edit after which no linear classifier can predict a given concept better than a constant guess, choosing, among all edits with that guarantee, the one that moves the representations least on average under a wide range of distance measures. It rests on the result that a representation is linearly guarded exactly when its class-conditional means are equal — that is, when the cross-covariance between the representation and the concept label is zero. Because the edit depends only on the data's means and covariances, it is cheap enough to run at every layer of a large network, which its authors call concept scrubbing; they used it to gauge how far language models depend on part-of-speech information and to cut gender bias in BERT embeddings.
We applied LEACE to every layer and perplexity rose sharply, so the model clearly relies on that feature.
Origin
Introduced by Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff and Stella Biderman in 'LEACE: Perfect linear concept erasure in closed form' (arXiv, June 2023; NeurIPS 2023), largely from EleutherAI.
Usage note
The guarantee is linear only: nonlinear classifiers may still detect the concept. Because LEACE removes only what a linear model could exploit, it typically damages other information less than iterative or adversarial methods.
Related
- concept erasureconcept erasure, n.: Removing a specific idea — such as a person's gender, or an artist's style — from what an AI model can represent or produce, so the model can no longer use or show it.
- iterative nullspace projectioniterative nullspace projection, n.: An early method for removing a trait such as gender from a model's internal representations by repeatedly training a simple detector for it and then flattening away whatever the detector relied on.
- directional ablationdirectional ablation, n.: Erasing one specific direction from an AI model's internal signals while it runs, so whatever that direction encodes — such as the tendency to refuse — can no longer influence what the model does.
- linear probelinear probe, n.: The simplest kind of probe: a straight-line classifier trained on an AI model's internal activity to test whether some piece of information can be read off directly.
- protected attributeprotected attribute, n.: A personal trait, such as race, sex, religion, age or disability, that the law says must not be used to treat people unfairly.
Sources
- Belrose, N., Schneider-Joseph, D., Ravfogel, S., Cotterell, R., Raff, E. & Biderman, S. (2023). LEACE: Perfect linear concept erasure in closed form. NeurIPS 2023. arXiv:2306.03819. arxiv.org/abs/2306.03819
Mentioned in
concept erasureconcept erasure, n.: Removing a specific idea — such as a person's gender, or an artist's style — from what an AI model can represent or produce, so the model can no longer use or show it., iterative nullspace projectioniterative nullspace projection, n.: An early method for removing a trait such as gender from a model's internal representations by repeatedly training a simple detector for it and then flattening away whatever the detector relied on.