Learning Dynamical Transport without Data

Overview

Join us as we explore advanced dynamical transport algorithms for generative modeling, entirely independent of data inputs. This exploration emphasizes the process of sampling from target distributions utilizing unnormalized log-likelihood functions.

Our insights can be directly applied to multifaceted domains such as physics, chemistry, and Bayesian inference. Dive into this innovative approach, available via YouTube, that merges artificial intelligence with core scientific disciplines.

Syllabus

• Introduction to Dynamical Transport
Overview of dynamical transport algorithms
Importance in generative modeling
• Sampling from Target Distributions
Unnormalized log-likelihood functions
Challenges of sampling without direct data
• Theoretical Foundations
Mathematical formulation of dynamical transport
Key principles of stochastic differential equations (SDEs)
Introduction to measure transport and transformation
• Methods and Techniques
Langevin dynamics
Hamiltonian Monte Carlo (HMC)
Normalizing flows
Score-based generative models
• Applications in Physics
Phase space sampling
Quantum systems and path integrals
• Applications in Chemistry
Molecular dynamics for reaction pathways
Importance sampling in chemical systems
• Applications in Bayesian Inference
Prior distribution sampling
Posterior estimation without data
• Computational Aspects
Numerical integration techniques
Efficient computation strategies
• Case Studies
Real-world examples from physics
Chemical systems simulations
Bayesian inference scenarios
• Conclusion
Recap of key concepts
Future perspectives in dynamical transport
• Project and Assessment
Design and implement a dynamical transport model for a chosen application
Evaluation based on accuracy, efficiency, and innovation

---

Subjects

Computer Science