Assignments
- Assignment #1, solution
Out Mon Jan 6, 2020 due Wed Jan 15, 2020
- Assignment #2, solution
Out Wed Jan 15, 2020 due Wed Jan 22, 2020
- Assignment #3, solution
Out Wed Jan 22, 2020 due Wed Jan 29, 2020
- Assignment #4, solution
Out Feb 17, 2020 due Feb 24, 2020
- Assignment #5
Out Mar 4, 2020 due Mar 11, 2020
Final Project
The final project is worth 40% of your final grade and can be done in teams of 2 or 3. Students should be putting in a substantial amount of effort into the project (about 4-5 weeks worth of work). The project should involve graphical models and will require a major programming effort, in the programming language of choice by the student team.
The project should simulate a research project. This means the project should have a research question or hypothesis that the project investigates. The team must also conduct experiments to test the hypothesis. I consider a project to be worth a B or C grade if you implement an existing algorithm from a paper. To get an A, you must extend the existing work in some "interesting" way that can be reasonably done in 3-4 weeks. The best case scenario is for your class project to become a conference or workshop paper.
The deliverables for the project are as follows:
- Before the project begins, you will submit a 1 page proposal describing your project.
- There will be a checkpoint 2 weeks into the project to make sure that you are making progress.
- A written report. The report should follow the structure of a research paper with the following sections: introduction, related work, methodology, results, discussion, and conclusion. Please use Latex or Word.
- Any code you used for the project. Please zip and submit to TEACH under the "project" submission.
- A 15 min oral presentation during exam week.
Projects are due March 18, 2020 at 23:59:59. Please slide the writeup under my door (or email the pdf) and submit the code via the TEACH handin.
Project Proposal
Your project proposal should be about 1 page and should have the following:
- The names and user IDs of the members of your group.
- 1-2 paragraphs describing what your project is about. In particular, describe what your main objective is and what experiments you will run to demonstrate it.
- A few sentences describing your deliverables for the midway checkpoint (roughly 2 weeks after your project starts)
Ideas for your final project
If you would like to look for papers related to probabilistic graphical models, take a look at the following conferences and journals:
Conferences: Uncertainty in AI (UAI), International Conference of Machine
Learning (ICML), Neural Information Processing Systems (NeurIPS, previously
called NIPS), AAAI conference on Artificial Intelligence (AAAI), International
Joint Conference on AI (IJCAI), AI and Statistics (AISTATS)
Journals: Journal of Machine Learning Research (JMLR), IEEE Transactions on
Pattern Analysis and Machine Intelligence (TPAMI), Journal of Artificial Intelligence Research (JAIR)
- Implement Latent Dirichlet Allocation.
If you decide to do this, I recommend implementing inference using Collapsed Gibbs Sampling. You should probably read Gibbs Sampling for the Uninitiated to get an understand of basic Gibbs Sampling on a topic model. Then, read this paper by Griffiths and Steyvers to understand Collapsed Gibbs Sampling.
In order to evaluate your model, you will need to implement a way to compute perplexity. I recommend the "Left-to-Right" method in this paper.
- N-mixture models are used in species distribution modeling to model counts of individuals of a particular species. You can implement extensions to a N-mixture model and apply it to eBird data.
- Fast learning of mixtures of tree models.
Learning with Mixtures of Trees by Marina Meila and Michael I. Jordan
- Explore extensions to the work on model averaging of Naive Bayes models.
References:
Exact model averaging with naive Bayesian classifiers by Denver Dash and Gregory F. Cooper
Model averaging for Naive Bayes models. by Ga Wu and Scott Sanner
- Do a "bakeoff" project. A bakeoff project is one that does an empirical comparison between multiple methods related to graphical models. For instance, you
could do a bakeoff between different recent approximate inference techniques (note that I said recent here). To get an A, you would need to do a thorough job
with the empirical comparison.
- Related to the bakeoff project, you could compare deep learning methods versus probabilistic graphical model methods for some task.
- If you are in both the probaiblistic graphical models and the deep learning
class, you can investigate the following papers that fall at the intersection
of probabilistic graphical models and deep learning:
- Reproducibility investigation: Take a recent paper on probabilistic
graphical models that has source code available and see if you can reproduce
the results. You would also need to propose a new set of experiments to
investigate some aspect of the approach that was not in the original paper.
- Explore recent approaches to Marginal MAP calculation. Note that you will
need to read about Marginal MAP in your textbook.
- Reinforcement learning: use a probabilistic graphical model to learn a model of the environment (e.g. "Learning and Querying Fast Generatie Models for Reinforcement Learning" by Buesing, Weer, Racaniere, Eslami, Rezende, Reichert, Viola, Besse, Gregor, Hassabis and Wie
stra")
- You may also get some ideas from Reinforcement Learning and Control as Probabilistic Inference: Tutorial and Review by Sergey Levine. This paper links RL with PGMs.
Resources
Bayesian network packages
Data Repositories
Project Grading
1. Project Checkpoint (5%) (due Feb 28, 2020 in class)
The project milestone involves handing in a 1 page report describing the milestones met along with any problems encountered and solutions for solving them. Please also submit a copy of your project proposal 1-pager.
2. Written Report (65%) (due March 18, 2020 at 23:59:59 pm PST)
The written report should be 4-8 pages using the standard templates for AAAI or ICML. The 8 pages is a max limit but references do not count as part of the 8 pages. Your written report should be like a research paper in which you empirically compare your approach against other competitive approaches and discuss the pros and cons. Your written report should have the following sections: Introduction, Related Work, Methodology, Results, Discussion, Conclusion, Bibliography.
When grading your written report, I will be evaluating it as a research paper. This means I will be evaluating it according to the following criteria:
- Novelty (5%): How original is this work? Are you simply re-impleme
nting a paper or are you trying out something that is different or unexplored?
- Motivation/Relevance to the course (5%): How is this project relat
ed to Probabilistic Graphical Models? What is the reason for investigating this
topic?
- Effort (20%): How much effort went into this project?
- Methodology (15%): What is the research hypothesis of your project?? How well do you investigate it?
- Results (15%) : Do the experimental results adequately test this hypothesis?
- Conclusions (5%): What conclusions can be drawn from these experiments? Do the experiments support the hypothesis?
You may submit the pdf for your written report through the TEACH handin system. Please submit your source code through TEACH or provide a Github link.
3. Oral Presentation (30%) (sign up on this Doodle link)
The oral presentation will be 15 mins with 5 mins for questions. Every teammust sign up for a project presentation during the finals week (March 16-20). Sign up for a 30 min slot (link to be provided). When signing up, please put down all the names of your group members. All presentations are via Zoom. All group members must be at the presentation unless I have permitted certain absences.
Back to home