I was giving a presentation for the final project on COMP790-142 Distributions and Collections in Machine Learning. Our project was about zero-shot learning. We basically did zero interesting thing on this project. But what I want to keep a memo for the presentation, because it is the first time I’m writing a transcription of all I should say, and practiced a little bit to catch the time limit.

The first thing to know is the requirements and audience of the talk. The advisor asked us to make only two slides and give a 5 min presentation. The audience is the (mostly I think) graduate students attending this class. So they should have basic understanding of machine learning, deep learning and some basic generative model, but may not know details about a specific area, which is zero-shot learning in our case.

I think a typical research presentation should spend at least $20\%$ of the time talking about the result. But in our case, we basically do not have many result to show. So generally I spend almost all the time I have talking about basics on zero-shot learning. Given the limitation of two slides, I start my pre with something like a thesis statement: We deal with generalized zero-shot learning problem with an improved generative model. Then I organized the pre by talking about the concept in the statement one by one:

  • What is zero-shot learning (and how to train)?
  • Why do we need generative model?
  • What is generalized zero-shot learning (and why previous method fail on this setting)?
  • How do we improve that?

Some lessons I learned from the experience:

  1. Before my talk I learned that the average person speaks at somewhere between 125 and 150 words per minute. But I think there might be some illustrations, gestures, pause etc., when talking with some slides, so I aimed for 100 words per minute. Actually it turns out that 100 is still a little tight for me…
  2. Be very familiar with your transcription. You should almost remember everything you write and should not even need to take the transcription in order to give a fluent talk. I am not familiar with some sentences in the transcription, and I was like ummm just at those places.
  3. Avoid using long sentence. Giving a talk is very unlike writing a paper. Long sentence is both difficult to read and to understand. I have a sentence in my transcription: In contrast, if you don’t use a distribution, you will have to, for example, map image and attribute features into points in the latent space, and conduct a nearest neighbor search to make the decision. I was almost running out of air after I said this…