Deep Probabilistic Programming Course @ DIKU

Last update: May 14, 2022

Related tags

Deep Learning deep-probprog-course

Overview

Syllabus

Part I - Introduction to Deep Probabilistic Programming

Week	Topic	Exercise	Links
1	Introduction to Bayesian Inference	Read Pattern Recognition and Machine Learning (PRML), Sections 1.1-1.3, 1.5-1.6 & 2-2.3.4 (inclusive ranges), Intro to Bayesian updating paper, and Pyro paper. Form up groups and ask a question for each chapter/paper you have read.	Pattern Recognition and Machine Learning Bayesian Updating Paper Pyro Paper
2	Variational Inference	Read the Variational Inference paper and Pyro tutorials on Stochastic Variational Inference (SVI). Ask three questions about them. Use Pyro’s Variational Inference support to implement the kidney cancer model. See worksheet and Bayesian Data Analysis 3rd Edition (BDA3) Section 2.7.	Variational Inference Paper Worksheet Bayesian Data Analysis Pyro SVI tutorial: Part I and Part II Pyro Website
3	Hamiltonian Monte Carlo	Read paper on Hamiltonian Monte Carlo and blog post on gradient-based Markov Chain Monte Carlo (MCMC). Look at the source code for Mini-MC. Ask a question each for HMC, the Mini-MC implementation and discrete variable marginalization. Implement Bayesian Change-point model in Pyro using NUTS.	Hamiltonian Monte Carlo Paper Gradient-based MCMC Mini-MC implementation Change-point model Pyro NUTS Example
4	Hidden Markov Models and Discrete Variables.	Read Paper on Hidden Markov Models and ask three questions about it. Read Pyro tutorials on Discrete Variables and Gaussian Mixture Models. Read Pyro Hidden Markov Model code example and describe in your own words what the different models do. Add amino acid prediction output to the TorusDBN HMM and show that the posterior predictive distribution of the amino acids matches the one found in data.	Hidden Markov Models Pyro Discrete Variables Tutorial Pyro Gaussian Mixture Model Tutorial Pyro Hidden Markov Model Example TorusDBN Optional: Epidemological Inference via HMC
5	Bayesian Regression Models	Read PRML Chapter 3 on Linear Models. Ask 3 questions about the chapter. Read the Pyro tutorials on Bayesian Regression. Solve the weather check exercise in the worksheet.	Pyro Bayesian Regression: Part I, Part II Worksheet
6	Variational Auto-Encoders	Read Variational Auto Encoders (VAE) foundations Chapters 1 & 2, and Pyro tutorial on VAE. Ask three questions about the paper and tutorial. Implement Frey Faces model from VAE paper in Pyro. Rely on the existing VAE implementation (see tutorial link).	Variational Auto Encoders Foundations Pyro Tutorial on VAE
7	Deep Generative Models	Read one of these papers: Interpretable Representation VAE, Causal Effect VAE, Deep Markov Model or DRAW (one paper per group). Try out the relevant Pyro or PyTorch implementation on your choice of relevant dataset which was not used in the paper. Make a small (10 minute) presentation about the paper and your experiences with the implementation.	Deep Markov Model Interpretable Representation VAE Causal Effect VAE DRAW

Part II - Deep Probabilistic Programming Project

The second part of the course concerns applying the techniques learned in the first part, as a project solving a practical problem. We have several types of projects depending on the interests of the student.

For those interested in boosting their CV and potentially getting a student job, we warmly recommend working with one of our industrial partners on a suitable probabilistic programming project. For those interested in working with a more academic-oriented project, we have different interesting problems in Computer Science and Biology.

Industrial Projects

Company	Area	Ideas
Relion	Shift-planning AI	Shift planning based on worker availability, historical sales data, weather and holidays. Employee satisfaction quantification based on previously assigned shifts. Employee shift assignment based on wishes and need
Paperflow	Invoice Recognition AI	Talk to us
Hypefactors	Media and Reputation Tracking AI	Talk to us
‹Your Company›	‹Your Area›	Interested in collaboration with our group? contact Ahmad Salim to hear more!

Academic Projects

Type	Description	Notes/Links
Computer Science	Implement a predictive scoring model for your favourite sports game, inspired by FiveThirtyEight.	FiveThirtyEight Methodology and Models
Computer Science	Implement a ranking system for your favourite video or board game, inspired by Microsoft TrueSkill.	Microsoft TrueSkill Model Can be implemented in Infer.NET using Expectation Propagation
Computer Science	Use Inference Compilation in PyProb to implement a CAPTCHA breaker or a Spaceship Generator	Inference Compilation and PyProb. You can use the experimental PyProb bindings for Java. CAPTCHA breaking with Oxford CAPTCHA Generator. Spaceship Generator
Computer Science	Implement asterisk corrector suggested by XKCD	XKCD #2337: Asterisk Correction
Computer Science	Implement an inference compilation based program-testing tool like QuickCheck or SmallCheck	Inference Compilation QuickCheck SmallCheck
Computer Science	Magic: The Gathering, Automated Deck Construction. Design a model that finds a good deck automatically based on correlations in existing deck design. Ideas like card substitution models could be integrated too.	Magic: The Gathering - Meta Site
Computer Science	Use probabilistic programming to explore ideas for solving Eternity II (No $2 million prize anymore, but still interesting from a math point of view)	Eternity II
Biology	Auto-Encoders or Deep Markov Models for Protein Folding	Deep Markov Model Pyro Deep Markov Model
Biology	Inference Compilation for Ancestral Reconstruction	Inference Compilation and PyProb. Talk to us for details.
Biology	Recurrent Causal Effect VAE for modelling mutations in proteins	Talk to us for details.

Recommendations

Sometimes sampling can be slow on the CPU for complex models, so try using Google Colab and GPUs and see if that provides an improvement.

Acknowledgements

This course has been developed by Thomas Hamelryck and Ahmad Salim Al-Sibahi. Thanks to Ola Rønning for suggesting the Variational Auto Encoders Foundations paper instead of Auto-Encoding Variational Bayes which we originally proposed to read on week 3. Thanks to Richard Michael for testing out the course and provide us with valuable feedback on the content!