[Colloquium] Xiaotian Duan Candidacy Exam/Mar 5, 2024

[Megan Woodward]

This is an announcement of Xiaotian Duan's Candidacy Exam.
===============================================
Candidate: Xiaotian Duan

Date: Tuesday, March 05, 2024

Time: 10 am CST

Location: JCL 298

Title: Retrieval-Based Parameter Activation for Large Language Model Knowledge Editing

Abstract: Knowledge editing in Large Language Models (LLMs) aims to make targeted changes to specific data points to fix inaccuracies or biases in these models without mistakenly altering other knowledge or skills. The research of knowledge editing focuses on three key aspects: generalization (the model’s ability to apply edited information across various contexts), locality (precise edits without affecting unrelated information), and scalability (maintaining performance efficiency and stability with increasing edits). We propose two novel contributions: (1) a new knowledge editing benchmark designed to overcome the limitations of existing benchmarks, which lack the material for fine-tuning methods and comprehensive evaluations, and (2) a novel external-memory-based approach for knowledge editing that utilizes an embedding model with a vector store. Preliminary results show that compact embedding models can effectively differentiate the edited and unrelated facts, maintaining high accuracy even when scaled to thousands of edits. Integrating this approach with parameter-efficient fine-tuning strategies, this research aims to address all three critical aspects of knowledge editing: generalization, locality, and scalability.

Advisors: Rick Stevens

Committee Members: Rick Stevens, Fangfang Xia, and Ben Zhao





-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20240304/006c8270/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: knowledge_editing_proposal.pdf
Type: application/pdf
Size: 821335 bytes
Desc: knowledge_editing_proposal.pdf
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20240304/006c8270/attachment-0001.pdf>