Abstract: Current AI training methods align models with human values only after their core capabilities have been established, resulting in models that are easily misaligned and lack deep-rooted value systems. We propose a paradigm shift from "model training" to "model raising", in which alignment is woven into a model's development from the start.

Abstract:

In this talk I'll develop the conceptual basis of generative AI, establishing a link between dynamical-systems models such as neural ODEs/SDEs and matching-based generative modelling. The first part focuses on deriving the continuous flow matching objective and relating it to diffusion, Schrödinger bridges, and dynamic optimal transport. The second part focuses on generative modelling on discrete state spaces, establishing a link between discrete denoising diffusion models and discrete flow models.

Estimators assuming missingness at random (MAR) can fail under missingness not at random (MNAR). Introducing complete auxiliary variables sometimes restores MAR by breaking dependence between analysis variables and missingness. However, if the auxiliaries are themselves incomplete, MAR typically remains violated.

Sign up on Eventbrite: https://medai-july2026.eventbrite.co.uk

Diffusion models are now the dominant approach for high-fidelity image generation, yet they remain highly vulnerable to privacy attacks, including reconstruction and membership inference attacks (e.g.,