Skip to content

Erasing concepts from neural representations with provable guarantees

License

Notifications You must be signed in to change notification settings

EleutherAI/concept-erasure

Repository files navigation

Least-Squares Concept Erasure (LEACE)

Concept erasure aims to remove specified features from a representation. It can be used to improve fairness (e.g. preventing a classifier from using gender or race) and interpretability (e.g. removing a concept to observe changes in model behavior). This is the repo for LEAst-squares Concept Erasure (LEACE), a closed-form method which provably prevents all linear classifiers from detecting a concept while inflicting the least possible damage to the representation. You can check out the paper here.

Installation

We require Python 3.10 or later. You can install the package from PyPI:

pip install concept-erasure

Usage

The two main classes in this repo are LeaceFitter and LeaceEraser.

  • LeaceFitter keeps track of the covariance and cross-covariance statistics needed to compute the LEACE erasure function. These statistics can be updated in an incremental fashion with LeaceFitter.update(). The erasure function is lazily computed when the .eraser property is accessed. This class uses O(d2) memory, where d is the dimensionality of the representation, so you may want to discard it after computing the erasure function.
  • LeaceEraser is a compact representation of the LEACE erasure function, using only O(dk) memory, where k is the number of classes in the concept you're trying to erase (or equivalently, the dimensionality of the concept if it's not categorical).

Batch usage

In most cases, you probably have a batch of feature vectors X and concept labels Z and want to erase the concept from X. The easiest way to do this is by using the LeaceEraser.fit() convenience method:

import torch
from sklearn.datasets import make_classification
from sklearn.linear_model import LogisticRegression

from concept_erasure import LeaceEraser

n, d, k = 2048, 128, 2

X, Y = make_classification(
    n_samples=n,
    n_features=d,
    n_classes=k,
    random_state=42,
)
X_t = torch.from_numpy(X)
Y_t = torch.from_numpy(Y)

# Logistic regression does learn something before concept erasure
real_lr = LogisticRegression(max_iter=1000).fit(X, Y)
beta = torch.from_numpy(real_lr.coef_)
assert beta.norm(p=torch.inf) > 0.1

eraser = LeaceEraser.fit(X_t, Y_t)
X_ = eraser(X_t)

# But learns nothing after
null_lr = LogisticRegression(max_iter=1000, tol=0.0).fit(X_.numpy(), Y)
beta = torch.from_numpy(null_lr.coef_)
assert beta.norm(p=torch.inf) < 1e-4

Streaming usage

If you have a stream of data, you can use LeaceFitter.update() to update the statistics. This is useful if you have a large dataset and want to avoid storing it all in memory.

from concept_erasure import LeaceFitter
from sklearn.datasets import make_classification
import torch

n, d, k = 2048, 128, 2

X, Y = make_classification(
    n_samples=n,
    n_features=d,
    n_classes=k,
    random_state=42,
)
X_t = torch.from_numpy(X)
Y_t = torch.from_numpy(Y)

fitter = LeaceFitter(d, 1, dtype=X_t.dtype)

# Compute cross-covariance matrix using batched updates
for x, y in zip(X_t.chunk(2), Y_t.chunk(2)):
    fitter.update(x, y)

# Erase the concept from the data
x_ = fitter.eraser(X_t[0])

Paper replication

Scripts used to generate the part-of-speech tags for the concept scrubbing experiments can be found in this repo. We plan to upload the tagged datasets to the HuggingFace Hub shortly.

Concept scrubbing

The concept scrubbing code is a bit messy right now, and will probably be refactored soon. We found it necessary to write bespoke implementations for different HuggingFace model families. So far we've implemented LLaMA and GPT-NeoX. These can be found in the concept_erasure.scrubbing submodule.

About

Erasing concepts from neural representations with provable guarantees

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages