/ir18

Inference and Representation (DS-GA-1005, CSCI-GA.2569), fall 18

Primary LanguageTeX

Inference and Representation (DS-GA-1005, CSCI-GA.2569)

Course staff:

Name E-mail
Instructor Joan Bruna bruna@cims.nyu.edu
TA Sanyam Kapoor sanyam@nyu.edu

Syllabus

This graduate level course presents fundamental tools of probabilistic graphical models, with an emphasis on designing and manipulating generative models, and performing inferential tasks when applied to various types of data.

We will study latent graphical models (Latent Dirichlet Allocation, Gaussian Processes), state-space models for (Kalman Filter, HMMs), Gibbs Models, Deep generative models (Variational autoencoders, GANs) covering both the methods (inference, sampling algorithms, learning, exponential families) and modeling applications to text, images and physics data.

Lecture Location

Tuesdays, 4:55-6:35pm, in 60 FA 110

[Recitation/Laboratory] (required for all students)

Mondays, 4:55-6:35pm in 60 FA 110

Office hours

JB: Tuesdays, 3:00pm-4:45pm. Location: 60 5th ave, 6th floor, room 612.

SK: TBA.

Grading

problem sets (40%) + midterm exam (25%) + final project (30%) + participation (5%).

Piazza

We will use Piazza to answer questions and post announcements about the course. Students' use of Piazza, particularly for adequately answering other students' questions, will contribute toward their participation grade.

Online recordings

Most of the lectures videos will be posted to NYU Classes. Note, however, that class attendance is required.

Lab Sessions

This semester, the lab sessions will feature the inverse curricula pioneered by C. Resnick in my previous class (see here and here for more details). The two topics where we will apply depth-first-learning are Normalizing Flows and Introduction to Model-based RL.

Schedule

Week Lecture Date Topic Reference Deliverables
1 9/4 Lec1 Introduction and Logistics. Inference Examples. Bayesian Networks. Slides Murphy Chapter 1 (optional; review for most)

Notes on Bayesian networks (Sec. 2.1)

Algorithm for d-separation (optional)
PS1, due 9/11
2 9/11 Guest Lecture: Rajesh Ranganath (NYU)
3 9/18 Lec2 Undirected Graphical Models. Markov Random Fields. Ising Model. Applications to Statistical Physics. Slides Notes on MRFs (Sec. 2.2-2.4)

Notes on exponential families

Notes on Hammersley-Clifford Theorem
PS2, due 9/25
4 9/25 Lec3 The Hammersley-Clifford Theorem. Belief Propagation. Slides Barber 27.1-27.3.1

Murphy Sec. 24.1-24.2.4

Introduction to Probabilistic Topic Models

Explore topic models of: politics over time, state-of-the-union addresses, Wikipedia
PS3, due 10/9 ipython notebook

Project Proposal, due 10/23
5 10/2 Lec4: BP (cont'd). Gibbs Sampling. PCA. Slides Elements of Statistical Learning, Ch.14

Finding Structure in Randomness (...), Halko, Martinsson, Tropp
6 10/9 No Lecture (legislative monday)
7 10/16 Midterm Exam
8 10/23 Lec5 PCA (cont'd). ICA. The EM algorithm Slides Graphical Models, Exponential Families and Variational INference, Chapter 3 Variational INference with Stochastic Search PS4, due 11/5
9 10/30 Lec6 EM (cont'd), MCMC Slides Variational Inference: A review for Statisticians, by Blei, McAuliffe, Kucukelbir AutoEncoding Variational Bayes (Kingma, Welling
10 11/6 Lec7 MCMC (cont'd), Variational Inference Slides Graphical Models, Exponential Families and Variational INference, Chapter 3 Variational INference with Stochastic Search PS5
11 11/13 Lec8 VI (cont'd) Variational Autoencoders Slides References on Slides
12 11/20 Lec9 VAE, Structured Output Prediction Slides PS6, due 12/8
13 11/27 Lec10 Structured Output Prediction (cont'd), EP, Unrolling inference with Neural Networks Slides references in slides, Expectation-Propagation Notes
14 12/4 Lec11 Deep Generative Models (1/2): Implicit Modeling Slides Geometrical Insights for Implicit Modeling, Bottou et al. and references in slides Project writeup, due 12/19.
15 12/11 Lec12 Deep Generative Models (2/2): Auto-regressive models. Open Problems Slides
16 12/18 Final Day Poster Presentations of Final Projects

Location: Center for Data Science, 60 5th ave, in the 7th floor open space

Bibliography

There is no required book. Assigned readings will come from freely-available online material.

Core Materials

Background on Probability and Optimization

Further Reading

Academic Honesty

We expect you to try solving each problem set on your own. However, when being stuck on a problem, we encourage you to collaborate with other students in the class, subject to the following rules:

  • You may discuss a problem with any student in this class, and work together on solving it. This can involve brainstorming and verbally discussing the problem, going together through possible solutions, but should not involve one student telling another a complete solution.
  • Once you solve the homework, you must write up your solutions on your own, without looking at other people's write-ups or giving your write-up to others.
  • In your solution for each problem, you must write down the names of any person with whom you discussed it. This will not affect your grade.
  • Do not consult solution manuals or other people's solutions from similar courses.

Late submission policy

During the semester you are allowed at most two extensions on the homework assignment. Each extension is for at most 48 hours and carries a penalty of 25% off your assignment.