The SAFR AI Lab at Harvard led by Professor Seth Neel and Professor Salil Vadhan invites applications for a 2 year post-doctoral research position focused on the following collection of topics:

• Machine Unlearning: The study of how to efficiently delete information from training models, without having to retrain the model completely. Initially motivated by privacy concerns, unlearning will find applications in things like LLM safety training, remedying copyright violations from generative models etc. We did some of the first work on machine unlearning, and recently have worked on unlearning in LLMs, with several ongoing projects. We are particularly interested in evaluation for non-convex models, and in connections between unlearning and model editing/approximation. Here is a recent interview Seth did with Axios Science on the topic.
• Data Attributions: The recent work in the Madry Lab (TRAK, Datamodels) has shown that meaningful data attribution (determining which training samples are most responsible for a given model output) in large models is possible, but the theory is far from complete. There are many tantalizing empirical questions that remain including how to accurately measure attribution, how to scale attribution to very large models, how to improve methods like TRAK, and connections to other areas like fairness, data subset selection, unlearning etc.
• Privacy Attacks on Generative Models: We’ve done a lot of recent work quantifying the privacy risks that come from training LLMs on sensitive data but questions remain. While white-box MIAs achieve high accuracy against pre-trained LLMs, work from us and others has shown that black-box attacks on LLMs are very weak relative to other types of models. Can we develop strong black-box MIAs or is there a fundamental barrier here?
• Other topics potentially of interest: detecting A.I. generated content, watermarking, copyright issues in generative models.

The selected candidate will be expected to lead research in methodological and applied research on these topics with a particular focus on generative models (e.g., GANs, VAEs, diffusion models, LLMs etc.). The ideal candidate is:

• Already working on these and related areas, and some ideas in the these subfields they want to pursue (this is what we will discuss in the interview)
• Wants to use this position as a springboard to a tenure-track assistant professorship at a great university or a competitive industry research position
• Is highly technical and is excited about implementing large scale deep learning experiments on the Harvard cluster both independently, and with a larger group of researchers. You have experience experimenting with LLMs (Llama-2, Pythia, OLMO etc) and/or vision models (ResNets, Stable Diffusion etc). Needless to say strong programming skills (Python, Pytorch) and deep knowledge of machine learning and its applications are required.
• Is proactive and comes to each meeting with an organized account of progress and attempted directions, questions or confusions, and new ideas.
• Is excited about working with a hands-on and highly motivated advisor
• Has demonstrably strong research skills, ideally, with publications in top venues in machine learning and/or top-tier interdisciplinary journals -- although this is not a hard requirement (e.g., ICML, NeurIPS, ICLR, KDD, AAAI, AI STATS, Nature/Science family of journals, PNAS).

Position

This position comes with competitive post-doc salary and excellent benefits. Through Professor Neel you will be affiliated to Harvard Business School, and to the Kempner Institute for the Study of AI, $\text{D}^3$ Institute (d3.harvard.edu), the ML Foundations Group (mlfoundations.org), and the Theory of Computation Group (https://toc.seas.harvard.edu/). You will have access to various Harvard computing clusters with A100 GPUs (HBS, CANON (SEAS), Kempner (for special projects)).

SEAS where our lab is located (highlighted in pink, views 😎)

The inside of SEAS where our lab is located