Project 2 - Anything deep
Due date: Dec 3 2018, 1:00am
In this project you’re free to do whatever you’d like, as long as it involves deep networks. To successfully complete the project, your report and presentation should be at the level of a workshop submission at a top tier computer vision, graphics or machine learning conference (e.g. NIPS, ICML, CVPR, ICCV/ECCV, Siggraph (Asia), ACL, EMNLP, …). You are allowed to use your current ongoing research. However you’re not allowed to use research that has already been completed and submitted. The project can also be a seriously executed “joke” (it does not need to be useful for science).
To help you decide if your project idea meets our expectations we have a 2 minute flash presentation (without slides) for each team on Oct 22. You can use this presentation to get early feedback on your idea.
Finally, be ambitious! It is fine to fail in class, as long as you try something interesting (well evaluated and executed negative results will as high of a score as positive results).
|Oct 10||Project 2 out|
|Oct 22||Project 2 flash presentations|
|Dec 3||Project 2 due|
A writeup. Preferably in latex using the CVPR 2018 stylesheet. It should contain:
- An intro and abstract outlining the method you chose and why it’s better than alternatives.
- A technical section, with a detailed overview of your method and any extensions.
- A thorough evaluation of any aspects of your method. Try to justify every choice you made either with a sound argument plus citation or with an experiment.
You do not need to submit your code for this assignment.
How to submit
Please submit a pdf of your writeup as a single file named
[UTEID]_[You name].pdf (e.g.
pk0000_PhilippKraehenbuehl.pdf) on canvas. If your working in a team name the file
[UTEID1]_[You name1]_[UTEID2]_[You name2].pdf.
- Keep the submission under 20MB. Otherwise, email the TA.
- We trust you to not doctor the results. If the numbers look too magical the class and course staff might ask you additional questions during or after your presentation.
Each team has 8 min to present (plus 2 min questions). Not every team member needs to speak, if you don’t want. A good presentation should spend:
- 2 min or less motivating your problem
- be very visual, maybe use an example
- 1 min or less stating the problem you’re solving
- be very specific here
- this should follow naturally from the motivation
- 2-3 min presenting your solution
- Convey the high level picture first
- You can go into the mathy details if you’d like (if you think it’s interesting), but you don’t need to.
- 2-3 min on results
- Show how your method works better (or not) than prior work
- Compare to alternatives (ablate your method)
This project counts 40% towards your final grade. The 40 points are split up as follows:
- 15pt Originality
- How novel is the idea compared to what we read in class, and other older publications
- How ambitious is the project?
- 15pt Writeup
- Technical section
- Thorough evaluation
- 10pt presentation
Late submission policy
The presentation cannot be late, you will loose 10pt if you do not have a presentation ready. You will lose 4pt for every day the submission is late. You will also not be eligible for the best submission score.
If you need inspiration here are a few project ideas (feel free to completely ignore them):
- Is there a correlation between training algorithms and generalization accuracy?
- Do we really need batch-norm? Why / why not?
- How much do convnets remember in what layer?
- How much can we learn from synthetic data (and transfer to real data)? Where do things fail?
- What do feature learning algorithm actually learn? How semantic are those features?
- Connections between deep reinforcement learning and adverserial losses