Erasure of Unaligned Attributes from Neural Representations

TL;DR

The paper introduces AMSAL, an algorithm that erases implicit attribute information from neural representations by iteratively assigning and projecting data, effectively reducing bias across multiple datasets.

Contribution

It presents a novel spectral method for removing implicit attribute information from neural representations, extending bias removal techniques beyond explicit attribute alignment.

Findings

Bias can be effectively reduced using AMSAL on various datasets.

The method works well with multiple guarded and protected attributes.

Limitations arise when there is strong entanglement between task and attribute information.

Abstract

We present the Assignment-Maximization Spectral Attribute removaL (AMSAL) algorithm, which erases information from neural representations when the information to be erased is implicit rather than directly being aligned to each input example. Our algorithm works by alternating between two steps. In one, it finds an assignment of the input representations to the information to be erased, and in the other, it creates projections of both the input representations and the information to be erased into a joint latent space. We test our algorithm on an extensive array of datasets, including a Twitter dataset with multiple guarded attributes, the BiasBios dataset and the BiasBench benchmark. The last benchmark includes four datasets with various types of protected attributes. Our results demonstrate that bias can often be removed in our setup. We also discuss the limitations of our approach when there is a strong entanglement between the main task and the information to be erased.