Q&A

Update as needed.

Week 1

• What about part of models as pre-trained from other works? If the paper itself uses the third party pertained model as part of their work?
  • It’s ok to use pre-trained models.
• Does a paper from my lab mean that the relevant fields ? Or just authors ?
  • Authors. For example, don’t select a paper your senior lab members proposed.
• Setup-python is a GitHub command?
  • See this.
• Does the coding need to be finished before presenting?
  • Yes. (Ok to update it later)
• Is one gtx2080 enough for this course?
  • It depends on your project, but it’s usually enough.
• How to register for AWS SageMaker Stuido Lab
  • Please submit your email address from this page https://studiolab.sagemaker.aws/requestAccount
  • You can find example projects here https://github.com/aws/studio-lab-examples
  • u-tokyo.ac.jp account is better (It will be approved on a priority basis)
• Is student with basic knowledge and decent experience about CV,ml and dl can handle most of the course content?
  • Yes.
• Before presentation, do I need to write “GitHub actions”?
  • Yes.
• More of the novelty and application of the paper, less of the formulas and principles?
  • Novelty itself is not important. Please pick up a paper you think has substantially impacted the field.
• If possible, I want the lecture of “Suitable papers”
  • Ok.
• It seems that UTAS registration is not updated on ITC-LMS, will it be okay to submit today’s survey later?
  • Please register both UTAS and ITC-LMS. You must be able to find a survey at ITC-LMS. If you cannot, let me know.
• Is an Arxiv paper, still under review at some conference ok?
  • Not so much recommended because several arXiv papers are simply wrong. But if you think the paper is worth explaining, it’s okay.

Week 2

• If a paper applies their method to two different tasks or domains like cv or audio, is it okay to choose just one for the implementation?
  • ok, up to you.
• Is just re-implementation enough for the presentation? I mean, sometimes, benchmarking/improvement might take time, or if it it doesn’t work for some reason, bug fixes etc. And by the second presentation we can present about benchmarking if possible. I mean some have more time to fix bugs and benchmark, some less.
  • If you think benchmarking/improvement is a big contribution, re-implementation is enough.
  • But for the 2nd round presentation, please focus PRs
• Must the paper we choose to represent be related to Deep Learning?
  • Absolutely no. Non-machine learning papers are strongly welcome.