Singular Learning Theory and Alignment Summit 2023

statistics
events
AI
Author
Published

2023-06-17

Abstract

There are surprisingly deep connections between algebraic geometry and statistical learning theory, which have been explored over the past few decades by Sumio Watanabe in Singular Learning Theory. This body of work is highly relevant to the problem of understanding the behaviour of large neural networks, thus providing a promising approach to AI alignment. The two-week summit will be held June 19 - July 1, 2023, with the first week of tutorials live-streamed from the Topos Institute. The second week features research talks and will be hosted at the Rose Garden Inn. The schedule and list of speakers can be found below. Please feel free to join any of the events! https://singularlearningtheory.com/

1 Introduction

There are surprisingly deep connections between algebraic geometry and statistical learning theory, which have been explored over the past few decades by Sumio Watanabe in Singular Learning Theory. This body of work is highly relevant to the problem of understanding the behaviour of large neural networks. In the Primer running over the first week of the SLT & Alignment summit, an introduction to the conceptual and technical foundations of SLT and associated statistical physics will be given, together with the outline of a promising research agenda for applying these ideas to problems in AI alignment. These tutorials will be live-streamed from the Topos Institute. The second week features research talks and will be hosted at Atlantis (2401 Piedmont Ave).

Picture from Sumio Watanabe

Picture from Sumio Watanabe

2 Event details

The schedule and list of speakers can be found above. Registration is free. Please join us for any of the events!

3 What is Singular Learning Theory?

It is a powerful method for analyzing the asymptotic behavior (as number of data samples go to infinity) of learning algorithms, especially for statistical models with hidden states such as mixture models, deep neural networks and partially observed Markov decision processes. Typically, we take a parametric statistical model and look at the algebraic/analytic variety of parameters whose model distribution minimizes the Kullback-Leibler divergence to some true distribution. By finding a change of parameters that resolve the singularities of this variety, we are able to derive many properties of learning algorithms for the statistical model, such as its generalization error. The hope is to use SLT in understanding what happens during learning, so that we can align the resulting AI with desirable traits.

4 What is AI Alignment?

AI (artificial intelligence) alignment is the general problem of getting an AI system to perform a desired task in a way that satisfies our preferences and ethical principles such as transparency and fairness. Here, we include the situation where the AI system (individually or collectively) is more intelligent than humans (individually or collectively) by some measure of intelligence. To this end, it seems robustly useful to understand the structure of knowledge and computation contained within the AI system and the process by which that knowledge is constructed. This will not solve AI alignment on its own, but significant progress on that problem will likely involve radical breakthroughs on this front.

Leaving a comment will set a cookie in your browser. For more information, see our cookies policy.