Speaker: Sam Gunn (MIT)

Title: Efficient deletion of data from trained AI models

Abstract: How does the choice of training data influence an 
AI model? This question is of central importance to 
interpretability, privacy, data attribution, and basic 
science. At its core is the data deletion problem: 
after a reasonable amount of precomputation, 
quickly predict how the model would behave if a 
given subset of training data had been deleted.
I will present a data deletion scheme capable of 
predicting model outputs to arbitrary precision. A 
bound on the error can be derived from a simple 
assumption about the stability of the AI model. In 
contrast to the assumptions made by prior work, 
ours appears to be fully compatible with deep 
learning. We believe that this method opens up a 
range of new possibilities for theory in AI. For 
instance, it gives the first ‘machine unlearning’ 
scheme with provable security in the deep learning 
setting.

Based on my recent work “How to sketch a learning 
algorithm” at https://arxiv.org/pdf/2604.07328.

Biography: Sam is a new CLE Moore Instructor at MIT. He was 
advised by Umesh Vazirani for his PhD at UC 
Berkeley. He is interested in studying the foundations 
of AI using ideas from cryptography and theoretical 
computer science.