Half time seminar by Mayank Nautiual: “Likelihood-free inference using Machine Learning”

Welcome to a half-time seminar in Scientific Computing presented by Mayank Nautiual. The seminar is held in English.

External reviewer: Ashkan Panahi, Chalmers University of Technology.

Abstract: Simulation-based inference (SBI) is a powerful approach for parameter estimation in scientific domains where likelihood functions are intractable but forward simulations are feasible. This seminar presents two complementary generative modelling approaches developed to accelerate SBI workflows.

The first approach, leverages conditional variational autoencoders (c-VAEs) to approximate complex posteriors efficiently using variational inference. We explore two variants of conditional VAEs — one that uses a learned prior conditioned on the observed data, and another that relies on a fixed standard Gaussian prior. The trade-offs between these approaches are evaluated in terms of modelling accuracy and computational efficiency across standard benchmark simulation tasks.

To address the limitations of VAE-based approaches—such as restrictive Gaussian assumptions, challenging training dynamics, and difficulty capturing complex multimodal distributions—we introduce ConDiSim, a conditional diffusion model for SBI. ConDiSim employs denoising diffusion probabilistic models (DDPMs) to approximate the posterior through iterative refinements, conditioned on observed data. This framework, enables richer generative capacity and more accurate mode coverage compared to VAEs.

We conclude the presentation with an outlook and future work.