Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers

TL;DR

Experiments show that C2B discovers biases beyond those of the original datasets and outperforms a recent state-of-the-art bias detection baseline that relies on task-specific annotations, being a promising first step toward addressing task-agnostic unsupervised bias detection.

Abstract

A person downloading a pre-trained model from the web should be aware of its biases. Existing approaches for bias identification rely on datasets containing labels for the task of interest, something that a non-expert may not have access to, or may not have the necessary resources to collect: this greatly limits the number of tasks where model biases can be identified. In this work, we present Classifier-to-Bias (C2B), the first bias discovery framework that works without access to any labeled data: it only relies on a textual description of the classification task to identify biases in the target classification model. This description is fed to a large language model to generate bias proposals and corresponding captions depicting biases together with task-specific target labels. A retrieval model collects images for those captions, which are then used to assess the accuracy of the model w.r.t. the given biases. C2B is training-free, does not require any annotations, has no constraints on the list of biases, and can be applied to any pre-trained model on any classification task. Experiments on two publicly available datasets show that C2B discovers biases beyond those of the original datasets and outperforms a recent state-of-the-art bias detection baseline that relies on task-specific annotations, being a promising first step toward addressing task-agnostic unsupervised bias detection.

Figures (23)

• Figure 1: We explore the novel task of bias discovery when we are only given a specific classification task and a pre-trained model. We propose Classifier-to-Bias (C2B), which automatically detects potential biases in the model, categorizes these biases, and assigns scores to each category.
• Figure 2: Overview of C2B. Given a specific task with an associated description and a pre-trained classification model, our approach leverages an LLM to identify potential candidate bias attributes and corresponding classes. These candidate biases are then used to prompt the LLM to generate captions, which are subsequently utilized by a retrieval module to collect a dataset of images for bias testing. This dataset enables evaluation of the target classification model, producing scores for each identified bias category.
• Figure 3: Examples of biases found by C2B for the high cheekbones target attribute on CelebA.
• Figure 4: Examples of biases found by C2B for the minibus class on ImageNet-X (ResNet50_V2).
• Figure 5: Examples of biases found by C2B for the birdhouse class on ImageNet-X (ViT_B_16_SWAG).