Odysea Session 8: Singular Learning Theory
About Event
Odysea is a free, flexible 8-week pilot for anyone in Southeast Asia curious about AI alignment. Inspired by Iliad's curriculum, it introduces the key ideas that alignment research draws on, including reinforcement learning, decision theory, training dynamics, and singular learning theory. Find out more at https://odysea-ai.org/.
📅 Event Details
Date: Thursday, November 12, 2026
Location: Online (Link provided upon registration)
Hosted By: RAIN-SG
💡 About This Session
Standard statistical learning theory assumes smoothly behaved, non-singular models—yet modern deep neural networks are inherently singular. Parameters are non-identifiable, the loss landscape is filled with complex geometries, and classical metrics often fail to explain why deep models generalize so well.
Enter Singular Learning Theory (SLT), pioneered by Sumio Watanabe.
SLT uses algebraic geometry and resolution of singularities to provide a rigorous mathematical framework for understanding learning, generalization, phase transitions, and capacity in singular statistical models. In this session, we will introduce the foundational concepts of SLT, explore its core analytical methodologies, and examine how it is being applied to decode complex AI systems today.
📊 Prerequisites
Multi-variable calculus and linear algebra
Basic familiarity with probability theory and statistical learning concepts
(Curiosity about algebraic geometry is a plus, but no prior background in algebraic geometry is required!)
Important notes
Registration is free. Please fill in all required fields. Registrations are subject to approval.
Food will not be provided. Feel free to grab a bite beforehand or bring a light snack, like bread or a sandwich.
Photography may take place during the event. By attending, you consent to being photographed.
We're committed to a respectful, inclusive space where curiosity and honest sharing can thrive, free from judgment.