# Lee Lab @ UW Madison

Research focus: Theory and algorithms for *foundation models*.

We sincerely appreciate the support provided by our sponsors: NSF, Amazon, and FuriosaAI.

## Openings

- [
**postdocs**] We are looking for a postdoc interested in the theoretical and algorithmic aspects of foundation models, particularly LLMs. If you want to work with us, please email me your CV and a short research statement. I strongly recommend reading our lab’s recent papers. - [
**PhD students**] We are looking for PhD students (ECE or CS) interested in the theoretical and algorithmic aspects of foundation models, particularly LLMs. If you want to work with us, please email me your CV and a short research statement. I strongly recommend reading our lab’s recent papers. - [
**MS students**] I am not currently looking for MS students. - [
**undergraduate students**] You might want to consider taking my class first. I teach various machine learning courses (ECE 532, 539, 561, 570, 761, …).

## News

- (July 2024)
**[COLM’24]**Can MLLMs Perform Text-to-Image In-Context Learning? is accepted to COLM’24! - (June 2024)
**[TMLR’24]**Mini-Batch Optimization of Contrastive Loss is accepted to TMLR! - (Apr. 2024)
**Amazon Research Awards**

Our group will develop principled approaches to prompt engineering through the information/coding-theoretic lens. Huge thanks to Amazon and my amazing collaborators and students! - (Mar. 2024)
**NSF CAREER Award**

Our group will develop a unified theory and new algorithms with provable guarantees for learning with frozen pretrained models, also known as foundation models. Huge thanks to NSF and my amazing collaborators and students!

## Recent Work on Foundation Models

Link | Topic/Type | Title or TLDR | Summary | Github |
---|---|---|---|---|

Arxiv’24 | LLM/Algorithm | Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data | Summary | Github |

COLM’24 | LLM/Algorithm | Can MLLMs Perform Text-to-Image In-Context Learning? | Summary | Github |

TMLR’24 | CLIP/Theory | Mini-Batch Optimization of Contrastive Loss | Summary | Github |

ICML’24 | LLM/Theory | Dual Operating Modes of In-Context Learning | Summary | Github |

ICML’24 | LLM/Algorithm | Can Mamba Learn How To Learn? A Comparative Study on In-Context Learning Tasks | Summary | Github |

UAI’24 | PEFT/Theory | Memorization Capacity for Additive Fine-Tuning with Small ReLU Networks | Summary | Github |

ICLR’24 | PEFT/Theory | The Expressive Power of Low-Rank Adaptation (LoRA) | Summary | Github |

ICLR’24 | LLM/Algorithm | Image Clustering Conditioned on Text Criteria | Summary | Github |

ICLR’24 | LLM/Algorithm | Teaching Arithmetic to a Small Transformer | Summary | Github |

ICLR’24 | LLM/Algorithm | A Looped-Transformer Architecture for Efficient Meta-learning | Summary | Github |

NeurIPS’23 | Diffusion/Algorithm | Reinforcement learning for improved text-to-image alignment | Summary | Github |

ICML’23 | LLM/Theory | Looped Transformers as Programmable Computers | Summary | Github |

ICML’23 | Diffusion/Algorithm | Reinforcement learning for faster DDPM sampling | Summary | Github |

NeurIPS’22 | LLM/Algorithm | LIFT: Language-Interfaced Fine-Tuning for Non-Language Machine Learning Tasks | Summary | Github |

NeurIPS’22 | Diffusion/Theory | Score-based Generative Modeling Secretly Minimizes the Wasserstein Distance | Summary | Github |

## Selected Talks on Foundation Models

- (Apr. 2024) The Johns Hopkins University CIS/MINDS seminar

Title: Theoretical Exploration of Foundation Model Adaptation Methods - (Mar. 2024) The 58th Annual Conference on Information Sciences and Systems @ Princeton University

Title: A Probabilistic Framework for Understanding In-Context Task Learning and Retrieval - (Feb. 2024) 2024 Information Theory and Applications Workshop

Title: The Expressive Power of Low-Rank Adaptation (LoRA) - (Feb. 2024) Foundations of Data Science - Virtual Talk Series @ UCSD/NSF TRIPODS Institute on Emerging CORE Methods in Data Science (EnCORE)

Title:*Theoretical Exploration of Foundation Model Adaptation Methods* - (Dec. 2023) CSP Seminar @ University of Michigan

Title:*Towards a Theoretical Understanding of Parameter-Efficient Fine-Tuning (and Beyond)* - (Nov. 2023) Efficient ML workshop @ Google Research New York

Title: The Expressive Power of Low-Rank Adaptation (LoRA)