Each student enrolled in the course will be required to present a lecture. Lectures will be presented by students in groups of 3-5. You can sign up for lectures here: https://tinyurl.com/yxn23usr. Each lecture will focus on a particular topic within deep probabilistic models, and is accompanied by one or more papers related to the topic. There are also additional papers related to each topic, which presenters may use for additional material. Note: you are not required to cover these additional papers. We (the TAs) will reach out to each group 1-2 weeks before the presentation date to check in, clarify technical details, and offer feedback on the presentation.
Each lecture should consist of the following:
- Approach: Describe the main idea of the topic in equations, words, and illustrations. Ideally, motivate the approach with a concrete example application/domain. Feel free to draw from resources outside of the assigned papers, such as other papers or textbooks, to present any relevant background material.
- Context: How does the topic fit into the broader framework of statistical modeling? Relate your topic to other ideas in deep probabilistic models, pointing out mathematical relationships, differing assumptions, etc. For instance, one could describe possible advantages and disadvantages of a modeling approach over others. In particular, relate your topic back to topics covered in previous lectures.
- The Paper(s): Describe the main points of the paper(s). What are the main claims? What empirical or theoretical evidence is provided to support the claims? Do not be afraid to be critical in your analysis; if you feel that there are conceptual gaps, un- or under-supported claims, missing baseline analyses, etc., then raise these points in your presentation.
- It is usually better to *not* follow the paper in linear order.
- Identify the key intuitions of the paper, and present those first, then dig deeper into the math as needed.
- Use figures & demos as often as possible.
- Presenting the full technical details is typically not desirable in a presentation.
- If you do present equations & formulas, explain carefully what each term means. If the formula has terms that you don't want to explain in a presentation, use a simplified formula.
We also encourage (but do not require) that you include:
- Tutorial: Briefly walk the class through an implementation of the algorithm, perhaps in a simplified, toy setting. This is often the best way to truly learn the topic. Feel free to draw on public repositories for your tutorial, ideally for inspiration. Describe important design choices in implementing the model, algorithm, or idea. We recommend that you use Jupyter notebook to present the tutorial.
Note that the previous points can be presented in various orders. We leave it up to each group to decide how to logically present the material.