BendVLM: Test-Time Debiasing of Vision-Language Embeddings

TL;DR

BendVLM introduces a nonlinear, input-specific debiasing method for vision-language model embeddings that avoids fine-tuning and is suitable for online, open-set applications.

Contribution

It presents BendVLM, a novel approach that performs input-tailored, fine-tuning-free debiasing of VLM embeddings, overcoming limitations of existing methods.

Findings

Effective reduction of societal biases in VLM embeddings.

Applicable to online, open-set tasks without prior input knowledge.

Outperforms linear debiasing methods in flexibility and effectiveness.

Abstract

Vision-language model (VLM) embeddings have been shown to encode biases present in their training data, such as societal biases that prescribe negative characteristics to members of various racial and gender identities. VLMs are being quickly adopted for a variety of tasks ranging from few-shot classification to text-guided image generation, making debiasing VLM embeddings crucial. Debiasing approaches that fine-tune the VLM often suffer from catastrophic forgetting. On the other hand, fine-tuning-free methods typically utilize a "one-size-fits-all" approach that assumes that correlation with the spurious attribute can be explained using a single linear direction across all possible inputs. In this work, we propose Bend-VLM, a nonlinear, fine-tuning-free approach for VLM embedding debiasing that tailors the debiasing operation to each unique input. This allows for a more flexible debiasing approach. Additionally, we do not require knowledge of the set of inputs a priori to inference time, making our method more appropriate for online, open-set tasks such as retrieval and text guided image generation.