How to Sketch a Learning Algorithm - Sam Gunn @videosfromIAS
How to Sketch a Learning Algorithm - Sam Gunn  @videosfromIAS
Uploaded May 2026 | Updated November 2026, 6 hours ago
Computer Science/Discrete Mathematics Seminar I
11:00am|Simonyi Hall 101 and Remote Access
Topic: How to Sketch a Learning Algorithm
Speaker: Sam Gunn
Affiliation: University of California, Berkeley
Date: May 28, 2026

How does the choice of training data influence an AI model? This question is of central importance to interpretability, privacy, 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 with arbitrarily small error ε and failure probability δ in the deep learning setting. The precomputation and prediction algorithms are only Õ(log(1/δ)/ε^2) factors slower than regular training and inference, respectively. The storage requirements are those of Õ(log(1/δ)/ε^2) models.

The proof is based on a simple assumption about the stability of training. In contrast to the assumptions made by prior work, stability appears to be fully compatible with learning powerful AI models. At a technical level, our work is based on a new method for locally sketching an arithmetic circuit by computing higher-order derivatives in random complex directions.

Based on my recent work: arxiv.org/pdf/2604.07328.
How to Sketch a Learning Algorithm - Sam GunnAdS3 Quantum Gravity via Finite-N Expansions - Wei LiSymplectophagy - Adi DicksteinFrom Fourier Restriction to Number Theory, Combinatorics, & Fractal Geometry III- Dominique MaldagueCreating Periodic Orbits of Reeb Vector Fields in Three Dimensions - Michael HutchingsFloer-Theoretic Entropy of Exact Symplectomorphisms - Joontae KimTaking the cosmic neutrino mass apart: geometry VS growth, and robustness to... - Emmanuel S SchaanIntersections and the Bezout Range - David UrbanikQuasinormal ringdown of quantum systems - Matthew Dodelson(Near-)extremal black holes beyond Einstein-Maxwell theory - Maciej KolanowskiInfinite-Order Lattice Anomalies and CPT - Salvatore PaceThe Role of Exposition in the Future of Mathematics - Grant Sanderson
Institute for Advanced Study |

How to Sketch a Learning Algorithm - Sam Gunn

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER