Skip to content

workshop.md

Latest commit

 

History

History
149 lines (102 loc) · 6.01 KB

workshop.md

File metadata and controls

149 lines (102 loc) · 6.01 KB

Workshop: Probabilistic Machine Learning

Instructor: Robert Osazuwa Ness
Author of Causal AI

Overview

This full-day workshop introduces intermediate and advanced practitioners to modern probabilistic machine learning, with a strong emphasis on uniting deep probabilistic modeling with Bayesian modeling techniques.

We focus on practical modeling workflows that combine statistical rigor with flexible, neural network–based architectures, using Pyro and PyMC. The workshop can be configured to:

  • Emphasize Pyro for deep probabilistic models, amortized inference, and close integration with PyTorch, or
  • Emphasize PyMC for expressive Bayesian modeling with solid NUTS/HMC workflows and integration with external deep learning components.

Participants will learn how probabilistic models support reasoning under uncertainty, scientific inference, and interpretable decision-making — and how to integrate these methods into real-world machine learning pipelines and research codebases.

The workshop blends hands-on demonstrations with conceptual foundations, making it suitable for both research scientists and industry data scientists who want to use probabilistic modeling as a core tool in their work.

Learning Objectives

By the end of this workshop, participants will be able to:

  • Build and critique Bayesian models using hierarchical structures, latent variables, and domain knowledge.
  • Select and apply inference algorithms such as NUTS, HMC, and stochastic variational inference (SVI).
  • Compare and combine computational frameworks Pyro and PyMC, choosing the right tool for the job.
  • Design and implement deep probabilistic models such as VAEs and normalizing flows.
  • Use probabilistic models for representation learning, forecasting, anomaly detection, causal inference, and simulation-based science.
  • Develop workflows for model criticism, sensitivity analysis, and communicating uncertainty.

Target Audience

  • Research scientists and ML engineers
  • Data scientists with experience in Bayesian statistics or ML
  • Researchers working on reasoning, uncertainty, generative modeling, or causal inference
  • Anyone using probabilistic programming in applied or academic settings

Attendees should be comfortable with Python, basic probability theory, and gradient-based model training.

Schedule (Full Day: 8:00–17:00)

8:00–8:30 — Introduction & Motivation

  • What probabilistic modeling is for
  • Why data scientists and researchers need uncertainty
  • When to choose probabilistic methods vs. classical ML or pure deep learning
  • How deep probabilistic models bridge structured reasoning and neural networks

8:30–9:45 — Bayesian Modeling Workflows

  • Model specification, prior elicitation, and prior push-forward checks
  • Posterior inference and posterior predictive checks
  • Workflow patterns from statistical and ML practice
  • Case Study: Bayesian model for structured prediction
  • How these workflows map onto Pyro and PyMC

9:45–10:00 — Break

10:00–11:30 — Inference Algorithms in Practice

  • Hamiltonian Monte Carlo and NUTS
  • Variational inference: SVI, amortization, reparameterization tricks
  • Trade-offs between accuracy, speed, latent variables, and scalability
  • Comparative demonstrations:
    • Using PyMC for NUTS/HMC workflows
    • Using Pyro for SVI and deep probabilistic models

11:30–12:30 — Lunch Break

12:30–13:45 — Hierarchical & Latent Variable Models

  • Partial pooling and hierarchical structure
  • Mixture models, clustering, and density modeling
  • Topic models, factor analysis, probabilistic embeddings
  • Case Study: Latent structure in behavioral or interaction data
  • Implementing hierarchical and latent-variable models in Pyro or PyMC

13:45–14:00 — Break

14:00–15:30 — Deep Probabilistic Models

  • Variational Autoencoders (VAEs)
  • Normalizing flows and expressive densities
  • Probabilistic neural networks and amortized inference
  • Hybrid models: combining neural nets with structured latent variables
  • Implementation focus:
    • Pyro for deep generative models and SVI
    • How PyMC can interoperate with external deep learning components
  • Example: representation learning and uncertainty-aware prediction with VAEs

15:30–15:45 — Break

15:45–17:00 — Causal Modeling, Decision-Making & Practitioner’s Playbook

  • Bayesian graphical models and structural assumptions
  • Identifiability and uncertainty-aware causal inference
  • Simulation-based causal analysis and counterfactual reasoning
  • Using deep probabilistic models in causal or “world-modeling” workflows
  • Practitioner’s playbook:
    • Choosing between Pyro and PyMC in practice
    • Communicating uncertainty to stakeholders
    • Common pitfalls and anti-patterns
    • Recommended reading and practice roadmap
  • Open Q&A

Software and Tools

This workshop uses:

  • Pyro for deep probabilistic models, amortized inference, and close integration with PyTorch.
  • PyMC for expressive Bayesian modeling with robust NUTS/HMC workflows and flexible integration with external deep learning models.

All examples are provided as runnable notebooks.

Suggested Background Reading

  • Statistical Rethinking — Richard McElreath
  • Bayesian Data Analysis — Gelman et al.
  • Official documentation for Pyro and PyMC
  • Selected research papers provided during the workshop

Instructor

Robert Osazuwa Ness is a researcher specializing in causal AI, generative modeling, and probabilistic programming. He is the author of Causal AI, has worked as an AI research scientist in both industry and academia, and is a contributor to leading open-source projects in probabilistic modeling.

Learn more at:
https://www.robertosazuwaness.com