During the last two semesters, I have been TA-ing the student seminar on Computer Vision and Deep Learning in our department. This includes guiding students on making their talks better. Many of the presentations, and frankly also my early presentations, suffered from similar drawbacks that can be improved easily. This post intends to help students tackle their first academic seminar presentations, and produce coherent, intriguing and enthusiastic talks.
All Eyes on Me
Arguably the most common feature of a poor presentation is showing too much information too fast. This includes anything from long sentences on your slides, complicated figures, cumbersome equations, and many more. That is not to say that texts/figures/equations should not be presented. It is how you present them that’s important.
When there’s text on the slide — the audience will read it. When there’s a complicated figure — the audience will try to understand what’s going on in there. When there’s an equation — the audience will try to grasp what it means. Reading text, understanding what’s in a figure, and grasping an equation takes time (it took you time too when you first saw them), and during this precious time, the audience stares at the slide instead of listening to you. You want to avoid such a scenario as much as possible.
Your goal is to make sure that the audience’s attention is directed at you at all times during your presentation.
Text on Slides
Some say, “you shouldn’t have text on your slides.” I agree and disagree. If you decide to put text on your slide (and want to keep the audience’s attention), make sure you read it out loud during your presentation. In some cases, it makes sense to put long text/sentences, like theorems, famous quotes or slides’ titles. In most cases, you’ll find out quickly that it’s not so necessary. If you’re not going to read it — don’t put it there.
Why do people still put text on slides? Some of us (especially less experienced speakers) fear that they will forget what to say during their talk and want some reminder. In this case, you can make a small list of reminders, 2–4 words noting what you are about to say. It is perfectly fine if your reminder is not a grammatically correct sentence. Avoid long sentences; use 2–4 words reminders.
Complicated Figures, Equations, or Anything Else
The same goes here; avoid showing a highly complex figure at once. Instead, start with some part of the figure, explain it, then reveal the next part and explain it, and so on.
With equations, the trick is to slowly walk the audience through each term in the equation, marking the relevant term that is being explained. This is good for guiding the audience to follow the relevant part of the equation at each point. But this simple trick also helps in “forcing” you to maintain a slow pace while presenting. It’s easy to talk too fast with all the adrenaline rush of a presentation and skip essential parts of the equation, so this trick is a speech-speed regularizer.
The Principle of Equal Speech-Slides Information Deltas
The above is neatly summarized with the following principle (shown in the figure below). Think of it, your presentation is actually two simultaneous presentations: one conveyed by your slides, the other conveyed by your speech. You want these two presentations to smoothly flow together, as balanced as possible.
At each “click” on your slides, new information is shown on the slides. Let’s mark by △(slides) the amount of information added following a “click”. Let’s mark by △(speech) the amount of information conveyed by your speech between two “clicks”.
A good presentation has △(slides)=△(speech) at each point in the presentation.
When too much information is added after one “click” (△(slides)>△(speech)) — the audience stares at the slides tryinig to figure out what’s going on in there, instead of listening to you. When they will (hopefully) finish staring and try to go back to listening to you, you’ll already be talking about something else. You probably lost most of your audience at this point :(
The other end (when △(speech)>△(slides)) means nothing has moved on your slides for too long. While some people are excellent speakers (and might not need a presentation at all), a presentation is a solid tool to keep your audience engaged. It’s a good rule of thumb to have something move on your slides every 10 seconds or so.
Here is how I imagine it graphically:
Each bump is a potential loss of audience attention. I always ask myself — Can I get this part of my presentation smoother?
How to Present a Paper
For most of our students (primarily first-year master’s students), this is the first time they are required to read and present a scientific paper. Here are some tips I found helpful on how to approach this:
Understand the Main Claims
Scientific papers, especially the applied ones, are full of technical details, experiments, hyperparameters choices, sometimes theorems with proofs, etc.
There’s no need to present every aspect that’s in the paper.
Start with understanding the central claims of the paper — What do the authors claim to have achieved? This might be a new state of the art on some benchmark, a new fundamental understanding of some aspect, a novel application, better image quality, etc. So you can first present the problem that the authors are tackling and then what they claim to have solved.
Presenting the Method
The next thing is understanding how they solved it (the model architecture/objective function used / main theorem). While you should understand the details, it doesn’t necessarily mean you should present all of them in your presentation. Only present the main elements (depending on how much time you dedicate to this paper). However, each introduced part should be explained sufficiently for the audience to understand it and be convinced that the authors’ claims are genuinely achieved.
“How deep should I delve into each part?” — this depends on the common knowledge of your audience. In our case, the audience should know linear algebra and the basics of deep learning, so it’s enough to only mention that “they used a slightly modified ResNet”. If the modification itself is not one of the main claims, there is no need to explain it further. If you want to present a theorem but only dedicate 2–3 minutes for it, you can explain the assumptions and the result of the theorem, and maybe give an intuition for the proof. But there is no need to fully prove all of it (unless that’s the central part of your presentation).
Choosing the main elements and presenting them fully is probably the most challenging part of making the presentation and requires lots of iterations.
Results — Convince the Audience (and Yourself) They Truly Did it
Ask yourself: Did the authors convinced you that their claims were valid? Did they back up their claims by experiments/theorem?
Here again, only show the main experiments/results (no need to present every table and every application). If possible, focus on visual results (look for the paper’s project page, it may have better results than the paper itself).
Unless very rare cases, avoid showing full benchmark tables (see the presentation slides/speech information delta principle above). Select only a few main essential numbers (and show them separately, no need to copy the original table it was in).
Be the Reviewer
Many students feel they should defend the paper. I even remember students presenting the paper in the plural form “we did this, we did that..”. Unless this is your paper, or you tried the method or ran the code (which is encouraged if possible), no need to take that position.
Your goal is different than the authors’. They mostly try to get their papers accepted to a conference/journal. Your goal is to understand their method and present it to the class. If you have a criticism: if something in their logic is not complete, if you think an essential number is missing in their experiments, if you think they could have done another experiment to validate their method, if you ran their code and the results ended up different than you had expected — all of these are great! Don’t hesitate to mention it in the relevant part of your presentation. To sum it all up:
Don’t take the position of the authors; take the position of the reviewer.
Dry Runs
Finally, it is hard to overestimate the importance of dry runs. Aim for at least 3–4 dry runs. Try to make them at least a few days before your talk — my point of view is that until the first dry run, when you think you’ve finished your work, you only finished 50% of the work.
A lot of things become clear during dry runs, you will spot many bumps in your presentation (which will require work to make them smoother), and probably after the 1–2 dry runs, the amount of text in your presentation will decrease dramatically (because you will already remember what you want to say).
Other helpful links
Other resources which I found both helpful and inspiring are this talk by Patrick Winston of MIT on how to give a good talk, and these notes by John E. McCarthy of Washington University on how to give a good colloquium.
I would like to thank Shir Amir for her helpful comments and proofreading
