Measuring Implicit Bias in Explicitly Unbiased Large Language Models

Large language models (LLMs) can pass explicit bias tests but still harbor implicit biases, similar to humans who endorse egalitarian beliefs yet exhibit subtle biases. Measuring such implicit biases can be a challenge: as LLMs become increasingly proprietary, it may not be possible to access their embeddings and apply existing bias measures; furthermore, implicit biases are primarily a concern if they affect the actual decisions that these systems make. We address both of these challenges by introducing two measures of bias inspired by psychology: LLM Implicit Association Test (IAT) Bias, which is a prompt-based method for revealing implicit bias; and LLM Decision Bias for detecting subtle discrimination in decision-making tasks. Using these measures, we found pervasive human-like stereotype biases in 6 LLMs across 4 social domains (race, gender, religion, health) and 21 categories (weapons, guilt, science, career among others). Our prompt-based measure of implicit bias correlates with embedding-based methods but better predicts downstream behaviors measured by LLM Decision Bias. This measure is based on asking the LLM to decide between individuals, motivated by psychological results indicating that relative not absolute evaluations are more related to implicit biases. Using prompt-based measures informed by psychology allows us to effectively expose nuanced biases and subtle discrimination in proprietary LLMs that do not show explicit bias on standard benchmarks.

Measuring machine learning harms from stereotypes: requires understanding who is being harmed by which errors in what ways

As machine learning applications proliferate, we need an understanding of their potential for harm. However, current fairness metrics are rarely grounded in human psychological experiences of harm. Drawing on the social psychology of stereotypes, we use a case study of gender stereotypes in image search to examine how people react to machine learning errors. First, we use survey studies to show that not all machine learning errors reflect stereotypes nor are equally harmful. Then, in experimental studies we randomly expose participants to stereotype-reinforcing, -violating, and -neutral machine learning errors. We find stereotype-reinforcing errors induce more experientially (i.e., subjectively) harmful experiences, while having minimal changes to cognitive beliefs, attitudes, or behaviors. This experiential harm impacts women more than men. However, certain stereotype-violating errors are more experientially harmful for men, potentially due to perceived threats to masculinity. We conclude that harm cannot be the sole guide in fairness mitigation, and propose a nuanced perspective depending on who is experiencing what harm and why.

Overcoming Bias in Pretrained Models by Manipulating the Finetuning Dataset

Transfer learning is beneficial by allowing the expressive features of models pretrained on large-scale datasets to be finetuned for the target task of smaller, more domain-specific datasets. However, there is a concern that these pretrained models may come with their own biases which would propagate into the finetuned model. In this work, we investigate bias when conceptualized as both spurious correlations between the target task and a sensitive attribute as well as underrepresentation of a particular group in the dataset. Under both notions of bias, we find that (1) models finetuned on top of pretrained models can indeed inherit their biases, but (2) this bias can be corrected for through relatively minor interventions to the finetuning dataset, and often with a negligible impact to performance. Our findings imply that careful curation of the finetuning dataset is important for reducing biases on a downstream task, and doing so can even compensate for bias in the pretrained model.

Gender Artifacts in Visual Datasets

Gender biases are known to exist within large-scale visual datasets and can be reflected or even amplified in downstream models. Many prior works have proposed methods for mitigating gender biases, often by attempting to remove gender expression information from images. To understand the feasibility and practicality of these approaches, we investigate what $\textit{gender artifacts}$ exist within large-scale visual datasets. We define a $\textit{gender artifact}$ as a visual cue that is correlated with gender, focusing specifically on those cues that are learnable by a modern image classifier and have an interpretable human corollary. Through our analyses, we find that gender artifacts are ubiquitous in the COCO and OpenImages datasets, occurring everywhere from low-level information (e.g., the mean value of the color channels) to the higher-level composition of the image (e.g., pose and location of people). Given the prevalence of gender artifacts, we claim that attempts to remove gender artifacts from such datasets are largely infeasible. Instead, the responsibility lies with researchers and practitioners to be aware that the distribution of images within datasets is highly gendered and hence develop methods which are robust to these distributional shifts across groups.

Measuring Representational Harms in Image Captioning

Previous work has largely considered the fairness of image captioning systems through the underspecified lens of "bias." In contrast, we present a set of techniques for measuring five types of representational harms, as well as the resulting measurements obtained for two of the most popular image captioning datasets using a state-of-the-art image captioning system. Our goal was not to audit this image captioning system, but rather to develop normatively grounded measurement techniques, in turn providing an opportunity to reflect on the many challenges involved. We propose multiple measurement techniques for each type of harm. We argue that by doing so, we are better able to capture the multi-faceted nature of each type of harm, in turn improving the (collective) validity of the resulting measurements. Throughout, we discuss the assumptions underlying our measurement approach and point out when they do not hold.

Towards Intersectionality in Machine Learning: Including More Identities, Handling Underrepresentation, and Performing Evaluation

Research in machine learning fairness has historically considered a single binary demographic attribute; however, the reality is of course far more complicated. In this work, we grapple with questions that arise along three stages of the machine learning pipeline when incorporating intersectionality as multiple demographic attributes: (1) which demographic attributes to include as dataset labels, (2) how to handle the progressively smaller size of subgroups during model training, and (3) how to move beyond existing evaluation metrics when benchmarking model fairness for more subgroups. For each question, we provide thorough empirical evaluation on tabular datasets derived from the US Census, and present constructive recommendations for the machine learning community. First, we advocate for supplementing domain knowledge with empirical validation when choosing which demographic attribute labels to train on, while always evaluating on the full set of demographic attributes. Second, we warn against using data imbalance techniques without considering their normative implications and suggest an alternative using the structure in the data. Third, we introduce new evaluation metrics which are more appropriate for the intersectional setting. Overall, we provide substantive suggestions on three necessary (albeit not sufficient!) considerations when incorporating intersectionality into machine learning.

Understanding and Evaluating Racial Biases in Image Captioning

Image captioning is an important task for benchmarking visual reasoning and for enabling accessibility for people with vision impairments. However, as in many machine learning settings, social biases can influence image captioning in undesirable ways. In this work, we study bias propagation pathways within image captioning, focusing specifically on the COCO dataset. Prior work has analyzed gender bias in captions using automatically-derived gender labels; here we examine racial and intersectional biases using manual annotations. Our first contribution is in annotating the perceived gender and skin color of 28,315 of the depicted people after obtaining IRB approval. Using these annotations, we compare racial biases present in both manual and automatically-generated image captions. We demonstrate differences in caption performance, sentiment, and word choice between images of lighter versus darker-skinned people. Further, we find the magnitude of these differences to be greater in modern captioning systems compared to older ones, thus leading to concerns that without proper consideration and mitigation these differences will only become increasingly prevalent. Code and data is available at https://princetonvisualai.github.io/imagecaptioning-bias .

Directional Bias Amplification

Mitigating bias in machine learning systems requires refining our understanding of bias propagation pathways: from societal structures to large-scale data to trained models to impact on society. In this work, we focus on one aspect of the problem, namely bias amplification: the tendency of models to amplify the biases present in the data they are trained on. A metric for measuring bias amplification was introduced in the seminal work by Zhao et al. (2017); however, as we demonstrate, this metric suffers from a number of shortcomings including conflating different types of bias amplification and failing to account for varying base rates of protected classes. We introduce and analyze a new, decoupled metric for measuring bias amplification, $\text{BiasAmp}_{\rightarrow}$ (Directional Bias Amplification). We thoroughly analyze and discuss both the technical assumptions and the normative implications of this metric. We provide suggestions about its measurement by cautioning against predicting sensitive attributes, encouraging the use of confidence intervals due to fluctuations in the fairness of models across runs, and discussing the limitations of what this metric captures. Throughout this paper, we work to provide an interrogative look at the technical measurement of bias amplification, guided by our normative ideas of what we want it to encompass.

ViBE: A Tool for Measuring and Mitigating Bias in Image Datasets

Machine learning models are known to perpetuate the biases present in the data, but oftentimes these biases aren't known until after the models are deployed. We present the Visual Bias Extraction (ViBE) Tool that assists in the investigation of a visual dataset, surfacing potential dataset biases along three dimensions: (1) object-based, (2) gender-based, and (3) geography-based. Object-based biases relate to things like size, context, or diversity of object representation in the dataset; gender-based metrics aim to reveal the stereotypical portrayal of people of different genders within the dataset, with future iterations of our tool extending the analysis to additional axes of identity; geography-based analysis considers the representation of different geographic locations. Our tool is designed to shed light on the dataset along these three axes, allowing both dataset creators and users to gain a better understanding of what exactly is portrayed in their dataset. The responsibility then lies with the tool user to determine which of the revealed biases may be problematic, taking into account the cultural and historical context, as this is difficult to determine automatically. Nevertheless, the tool also provides actionable insights that may be helpful for mitigating the revealed concerns. Overall, our work allows for the machine learning bias problem to be addressed early in the pipeline at the dataset stage. ViBE is available at https://github.com/princetonvisualai/vibe-tool.

Learning Robotic Manipulation through Visual Planning and Acting

Planning for robotic manipulation requires reasoning about the changes a robot can affect on objects. When such interactions can be modelled analytically, as in domains with rigid objects, efficient planning algorithms exist. However, in both domestic and industrial domains, the objects of interest can be soft, or deformable, and hard to model analytically. For such cases, we posit that a data-driven modelling approach is more suitable. In recent years, progress in deep generative models has produced methods that learn to `imagine' plausible images from data. Building on the recent Causal InfoGAN generative model, in this work we learn to imagine goal-directed object manipulation directly from raw image data of self-supervised interaction of the robot with the object. After learning, given a goal observation of the system, our model can generate an imagined plan -- a sequence of images that transition the object into the desired goal. To execute the plan, we use it as a reference trajectory to track with a visual servoing controller, which we also learn from the data as an inverse dynamics model. In a simulated manipulation task, we show that separating the problem into visual planning and visual tracking control is more sample efficient and more interpretable than alternative data-driven approaches. We further demonstrate our approach on learning to imagine and execute in 3 environments, the final of which is deformable rope manipulation on a PR2 robot.