Skip to content

Logistics

Inspiration Thesis

  • Neural Density Estimation and Likelihood-Free Inference - Papamakarios, 2019
  • Broadening the Scope of Gaussian Processes for Large-Scale Learning - Cutajar, 2019
  • Uncertainty in Deep Learning - Gal (2016)
  • Bringing Models to the Domain: Deploying Gaussian Processes in the Biological Sciences - Zwießele (2017)
  • Model-Based Understanding of facial expressions - Sauer (2013)
  • From Dependence to Causation - Lopez-Paz (2016)
  • Understanding Random Forests: From Theory to Practice - Louppe (2014)

Inspiring Talks

Story Principle

There is typically a 3-step rule for telling a story especially when you want people to learn something from it. We're stupid, we need repitition. The repeat rule of 3 is what I'll use to structure the thesis.

  1. Tell them what you're going to tell them
    • Key Ideas
  2. Tell them
    • Data Representation
    • Information
  3. Tell them what you told them
    • Discussion
    • Conclusion

Writing Principles (SEED)

I will try to follow this principle for writing. brief and succinct. I want my statements to really explore and develop one single idea before moving on to the next idea. So in order to keep things in control, I will follow the SEED principle.

  1. Statement
  2. Explanation
  3. Evidence
  4. Development
Resources * Writing Science: How to write papers that get cited and proposals that get funded - Joshua Schimel (2011)

Philosophy

Explanations

If you can't explain it simply, then you don't understand it enough - Einstein

Code Snippets

Talk is cheap. Show me the code - Linus Torvald

Lab Notebooks

OK. But how does it work in practice? AutoDiff all the things!

Sleeper Theorems

I'll leave it up to the user as an exercise | It's easy to show that | as seen in [1]

Parting Words

I like to leave parting words. I see this in a blog by Matthew Rocklin all of the time

  1. What I did I do
  2. What I did I not do
  3. What I could have done better
  4. What will I do in the future?

Another way to look at it is to have strengths and limitations in my discussions and conclusions. In my highschool, I typically had to do 5 of each; sometimes 10 of each if we behaved badly in class. It was tough but it made us think critically and reflect upon our work.

Details

Often times, we have many details that I think is important to know but not necessarily important to tell the story. Theorems, derivations, side notes, are very important but sometimes I think we can omit them from the main body of the text. I will create a few tabs that will hide some of these details I deem irrelevant for the story.

Details These are simply details that I feel are side notes.
Code Code examples that maybe tell you how to algorithmically do something. They might also feature snippets of some practical modifcations that may occur in the field.
Proof I am not a fan of proofs being in the main body of the text (unless the text is about proofs). In ML, often this is not necessary except for a theoretical paper. In applied settings, we only need main equations and the rest of the details can go in the appendix.
Resources Extra links where a better explanation can be given.

Reproducibility

  1. Repeatability: Same team, same experimental setup
  2. Replicability: different team, same experimental setup
  3. Reproducibility: different team, different experimental set up
Resources **Source**: Association for Computing Machinery (2016)