DME Paper Presentation

You will present a recent research paper in machine learning or data mining as part of the course. You will present the papers in groups, and there will be a class discussion afterwards.

All students are expected to have read the papers before class, although obviously not to the same amount of detail as the presenters. There will be questions on the exam about the paper presentations.

What to Talk About

The goal of the presentation is to explain the key points of the paper to the other students in the class. The mark will be based mainly on how well you understand the material, but also in part on the quality of the presentation itself.

Your presentation should help your classmates understand the paper. Your presentation should answer questions like

What problem is the paper trying to solve?
What was the state of the art before the paper was published?
How does the paper solve this problem?
What is the main contribution of this paper?
How did the authors evaluate their work?

In addition, during both presentation and the discussion, we want to evaluate the work, i.e., decide how good it is, and whether we would really use it in practice. Related to this point, here are some questions that you should consider either addressing during your presentation or bringing up during the discussion:

Based on the topics that you know from PMR, MLPR, and IAML, how might you have approached this problem? (This might be a good question for the discussion, e.g., Why didn't the authors just use k-nearest neighbour?)
Is their approach practical?
Would their technique apply to other similar problems? How general is it?
What are the limitations of their approach?

I encourage you to use the data projector to show slides from a laptop. You should plan on presenting the slides from my laptop. You must send me your slides on the day before your presentation.

You should feel free to use diagrams and equations from the paper as necessary in your slides.

Advice for Reading Papers

First, some advice about reading research papers. I endorse this advice. A succinct way of putting this is: If you really understand the paper, you should be able to summarize it, preferably in a different way than the authors did.

Reading in depth. The level of understanding of a research paper required to give a presentation on it is deeper than you may be used to. You will probably need to read the paper more than once, and key bits you may need to read several times over. You may want to try your own examples on the techniques described. As you prepare your talk you will discover that there are details that you did not quite grasp on first reading and you will have to read them up again. I would expect reading to this depth to require several hours at least.

Read critically. If you were skeptical or critical of some aspects of the paper, please do include this in your talk, as it makes for interesting discussion.

Advice for Giving Talks

First, Sharon Goldwater's advice is good. I would add:

Designing a talk. In 25 minutes there is no time for padding. Identify the main point(s) you want to convey and go as directly to it/them as quickly as possible. Reiterate this point(s) in title, introduction and conclusion. Give only the background information that is vital to understand the main point(s). Keep a mental model of the state of your audience's understanding: what have you told them so far? did they understand it? what do they need to know to understand the next point?

Preparing slides. Make sure the slides are visible from the back of the room. I prefer to make my slides as visual as possible, avoiding words where I can. It is perfectly OK to copy and paste graphics from the paper. When you use text, use a big font size. Do not write entire sentences on a slide. Keep the information content of each slide small. Use short noun phrases as slogans for each point of the argument. Avoid using too many slides. Avoid using too many words on a slide. When you use figures and graphs, explain what each part of the figure means. Explain what the axes on the graph are. If one of the axes is performance, are large numbers good or bad? (e.g., Are you plotting accuracy or error?)

Prepare your slides well in advance and get feedback on them. Try them out on the projector and walk to the back of the room to see what they look like. Practice giving your talk several times; 6 times is not too many.

Tips on speaking. Take your time. Most people talk too fast rather than too slow. Expect that your audience has very little memory, e.g., if you define some function on slide 2 and then don't mention it again until slide 10, don't expect them to remember it.

Getting started. It can be hard to get started giving a talk. A lot of people stumble through the first few words before they start. To avoid this, I usually memorize the first sentence of my talk, which is a short one sentence summary of the talk, e.g., "In today's lecture, I'm going to talk about probabilistic models for documents." It is counterproductive to attempt to memorize your entire talk.

Acknowledgement Thanks to Alan Bundy for permission to use material from his AI Research Methodology web pages.

This page is maintained by Amos Storkey

Home : Teaching : Courses : Dme