A Framework for Responsible Deployment of Large Language Models 

Abstract: The recent explosion in the capabilities of large language models has led to a wave of interest in how best to prompt a model to perform a given task. While it may be tempting to simply choose a prompt based on average performance on a validation set, this can lead to a deployment where unexpectedly poor responses are generated, especially for the worst-off users. To mitigate this prospect, I will describe a lightweight framework for selecting a prompt based on rigorous upper bounds on families of informative risk measures. This framework facilitates methods for producing bounds on a diverse set of metrics, including quantities that measure worst-case responses and disparities in generation quality across the population of users. Experiments on applications such as open-ended chat, medical question summarization, and code generation highlight how such a framework can foster responsible deployment by reducing the risk of the worst outcomes. I will also show that the same framework can be used to control several fairness measures, in a variety of applications, including medical imaging and film recommendation.

At the Intersection of Algorithmic Fairness and Causal Representation Learning

Abstract: At the intersection of algorithmic fairness and causal representation learning, an eminent question is whether we would like to simply impose fairness constraints, or it would also be fruitful to work with the data-generating process and to achieve the fairness goal. In the first part of the talk, we start by considering the attainability of a commonly-used group-level fairness notion in the static setting. We highlight the importance of investigating the theoretical possibility of actually achieving specific fairness criteria. Taking the temporal axis into consideration, we present Tier-Balancing notion of long-term and dynamic fairness, encapsulating both observed and latent variables in the pursuit of long-term fairness goals. In order to precisely capture fairness, or neutralness, in a principled way, the underlying data-generating process itself plays a crucial role. Hence, in the second part of the talk, we will see how to achieve causal representation learning, which reveals hidden causal variables and causal relations, and how one can further enhance fairness pursuits by decoupling objectionable data-generating components from their neutral counterparts.

Uncovering Hidden Bias: Auditing Language Models with a Social Stigma Lens

Abstract: Auditing LMs for unwanted social bias is challenging, not only due to their opaque behavior, but also due to the multidisciplinary nature of the work. This talk shows how we can learn from the social sciences domain and go beyond gender/race/religion in auditing language model outputs. Using a list of 93 US-centric stigmas documented in the social sciences literature, we curated a Q&A dataset - SocialStigmaQA - comprising 10K prompts, with a variety of prompt styles. We analyze the generated output of two open-source language models (Flan-T5 and Flan-UL2) and find that the proportion of socially biased answers for the Q&A task ranges from 45% to 59% across a variety of decoding strategies and prompting styles. We discover that the deliberate design of the templates in our benchmark (e.g., adding biasing text to the prompt or varying the answer that indicates bias) impacts the model tendencies to generate socially biased output. Through manual inspection of the generated chain-of-thought output, we identify a wide range of issues that range from subtle bias to lack of reasoning. 

This talk will offer a sneak-peak of our upcoming AAAI 2024 paper on SocialStigmaQA

Performativity and Power in Prediction

Abstract: Algorithmic predictions inform decisions, incentivize strategic actions, and motivate precautionary measures. As such, predictions used in societal systems not only describe the population they aim to predict, but they have the power to change it; a prevalent phenomenon often neglected in theories and practices of machine learning. In this talk, I will start by introducing the calculus of performative prediction, that conceptualizes this phenomenon by allowing the predictive model to influence the distribution over future data. This dynamic perspective on prediction elucidates new solution concepts, optimization challenges, and brings forth interesting connections to concepts from causality and game theory that I will discuss. Moreover, it allows us to articulate two mechanisms fundamental to prediction, learning and steering, making explicit the important role of power in prediction, and giving us a means to measure it. I will end my talk on a discussion of collective action as a strategy to resist the power of digital platforms. 

In light of the workshop, I hope my talk covers tools and concepts that will be useful for monitoring, assessing, and designing fair algorithmic systems.