Reading group dedicated to learning theoretical and practical foundations of GenAI, allowing us to integrate it with own research
When: Wednesdays at 11 am.
When: AI Institute meeting room.
Zoom link Deprecated, planned to be cancelled, (only if you really cannot make it in person): https://mit.zoom.us/j/97422870792.
Recording: Deprecated, planned to be cancelled, generally not recorded, ask the next presenter to record if you really cannot make it.
Important: Zoom link and recording should not be a regular excuse not to attend, and will soon become impossible.
(Exceptional permission to use zoom can be made for very long-term attendants)
-
We pick papers of general interest (i.e., no heavy neuro or molecular biology papers unless they truly introduce a general idea)
-
Everyone reads the paper beforehand
-
One person presents with slides
-
Everyone asks questions to understand the paper
-
We break up into groups of 2-3 and collect subjective perspectives on the topic
-
We reconvene to share ideas and identify promising research directions
-
Communication works via GitHub issues (one issue per paper) and discussions (everything else). Make sure to enable notifications for these two parts in the repository
-
The presenter should moderate, i.e. steer the discussion (avoid derailing)
-
Joel (@SeasonsOfTheSun) is permanent auxiliary moderator
-
Outreach/Recruiting for the club is currently only word-of-mouth, we will reevaluate in the future
- ~3 month blocks
- Theory basics
- Seminal papers of the topic
- Applications
- Managed democracy
- Monthly discussions for feedback/ideas
- GH discussion is set up in the beginning of each month and addressed at the end of the month in our meeting (15mins)
- Monthly discussions for feedback/ideas
- Have eventually brainstorming sessions / short project presentations (RELATED TO PREVIOUSLY INTRODUCED TOPICS)
- GH Discussion post to group-brainstorm the next block
- Orga team meets to consolidate the suggestions
- Publishing (GH post, Website, Poster, …)
- Gathering Materials for Course
- Joel will be responsible for this
- Paper and/or code (or both) (try to get a mixture of session kinds)
- 5 min presenter-feedback
- 5 min organisation
- Which next paper
- Who does next paper
Joel:
- Auxiliary chair,
- Curating materials for making a lecture course,
- Contribute to MedUni course registration.
Moritz:
- General GitHub Orga.
Albert:
- Maintain monthly GH discussion feedback,
- Contribute to MedUni course registration.
- Focus on the paper that was agreed on and take enough time to prepare adequately (e.g. 2+ hours every day in the week before the meeting).
- Say what you will say. Have an up-front slide describing the take-home messages and why the paper should be worth remembering.
- Link the paper to other papers (mathematically and conceptually), especially those that have already been covered in previous sessions.
- Either explain something properly or leave it out. This is especially true for mathematical innovations. A formula will be forgotten, but explaining the core idea and, in many cases, the proof can be very insightful.
- Examples, examples. Examples are essential to understanding any theory, but in ML, this is twice as true because experiments usually indicate the type of data the algorithm is designed to work on.
- Say what you have said. Conclude with a slide that describes how the innovations fit in the bigger picture.
- Raise points of discussion. If you have found something difficult to understand, chances are those bits are worth discussing.
- 5 mins feedback in the end
- Optional but encouraged (Joel will ask the presenter before the presentation and then enforce)
- Purpose: Training for both (give feedback, receive feedback, improve presentation style)
- Feedback must be positive, encouraging and constructive (e.g. sandwich technique)
- Biscuits prize each month for best presentation (selected by Joel based on presenter feedbacks)
- Mirror screen -> PowerPoint -> Draw in power point