Debiasing Large Vision-Language Models by Ablating Protected Attribute Representations
Published in NeurIPS 2024 SafeGenAI Workshop, 2024
CV
Education
DPhil University of Oxford, 2022
Supervisors: Andrew C. Eggers, Raymond Duch
Wrote doctoral thesis on applications of RL and NLP in political campaigns:
- Conducted human subjects research demonstrating that offline RL could be used to improve the effect of targeted political advertisements.
- Designed and trained multimodal transformer-based encoder for generating image-and-text embeddings from political Facebook ads.
- Compared structured and unstructured NLP tools for document retrieval of economic narratives in newspaper corpora.
Teaching: Python, statistics and HPCs to social scientists.
MSc University of Oxford, 2019
Supervisor: Andrew C. Eggers
Focused on quantitative/statistical methods for causal inference and NLP.
BA University of Oxford, 2015
Read Philosophy, Politics and Economics at Merton College, Oxford.
Employment
AI Research Scientist, Intel Labs, USA (2024-present)
Staff scientist with Multimodal Cognitive AI group under PI Vasudev Lal.
Research Highlights:
- First author on two new projects in first five months.
- Received Outstanding Paper Award at ACL 2024 for work on measuring latent values and opinions in LLMs.
- Won the Scholar Award for highest academic output in first year at Intel Labs.
Engineering Highlights:
- Ported cutting-edge multimodal models to new architectures.
- First team to execute foundation model training at 1K Gaudi 2 card scale in Intel Tiber AI Cloud.
Postdoctoral Research Associate, Princeton University, NJ, USA (2022-2024)
Postdoctoral researcher at Data Driven Social Science Initiative working with Professor Brandon Stewart.
- Published statistical framework for using LLMs for data annotation without introducing bias.
- Gave workshops on LLMs and intro deep learning to generalist audiences.
- Developed LLM agent-based tool for improving survey research.
Predoctoral Research Fellow, University College London, London, UK (2021-2022)
Predoctoral researcher on UKRI-funded MENMOPE project led by Professor Lucy Barnes.
Quantitative Analyst, SBI Group, Tokyo, Japan (2016-2018)
Data scientist/analyst at SBI Japannext, Japan’s foremost PTS.
- Automated reporting for stock market clients from terabyte-scale trading engine data.
- Used ML/econometrics to provide business intelligence for internal and external stakeholders.
Skills
Experienced with following:
- Languages:
Python,R,bash,SQL - Model development with
pytorch,transformers - Distributed training with
torch.distributed,deepspeed,FSDP - Experienced with
docker,kubernetesandSLURM
Publications
Steering Large Language Models to Evaluate and Amplify Creativity
Published in NeurIPS 2024 Creativity and Generative AI, 2024
We use model steering to measure and enhance creative properties of LLMs.
Exploring Vision Transformers for Early Detection of Climate Change Signals
Published in NeurIPS 2024 Tackling Climate Change with Machine Learning, 2024
AutoPersuade: A Framework for Evaluating and Explaining Persuasive Arguments
Published in EMNLP 2024 (Main), 2024
We introduce a three-part framework for constructing persuasive messages.
Why do LLaVA Vision-Language Models Reply to Images in English?
Published in EMNLP 2024 (Findings), 2024
Including an image in the query to a LLaVA-style VLM significantly increases the likelihood of the model returning an English response, regardless of the language of the query.
LLaVA-Gemma: Accelerating Multimodal Foundation Models with a Compact Language Model
Published in CVPR 2024 Multimodal Foundation Models Workshop, 2024
We train a suite of multimodal foundation models (MMFM) using the popular LLaVA framework with the recently released Gemma family of large language models (LLMs).
Political Compass or Spinning Arrow? Towards More Meaningful Evaluations for Values and Opinions in Large Language Models
Published in ACL 2024 (Main), 2024
We evaluate the latent political biases of LLMs, and show that models give substantively different answers when not forced; that answers change depending on how models are forced; and that answers lack paraphrase robustness.
Using Imperfect Surrogates for Downstream Inference: Design-based Supervised Learning for Social Science Applications of Large Language Models
Published in NeurIPS 2023 (Main), 2023
We present a new algorithm for using imperfect annotation surrogates for downstream statistical analyses while guaranteeing statistical properties—like asymptotic unbiasedness and proper uncertainty quantification—which are fundamental to Computational Social Science research.