Machine Learning Study Group
Welcome!
We
start at 4:00 (or as soon as the weekly seminar is over and people have
logged in). Zoom automatically ends the meeting after 40 minutes. New
members are always welcome and anyone can join. You may attend any or
all sessions. Also feel free to ask to be
removed from the invite list at any time, as we have no wish to send
unneeded emails which we all certainly get too many of.
Agenda & Minutes
- Welcome to the 63rd meeting, May 5, 2023.
- Status updates/news/inputs/comments
- Student and project updates
- ?
- Any other updates/news/inputs or comments?
- We read https://openai.com/research/gpt-4 about GPT-4 up to the section "Risks & mitigations" We could proceed from there.
- We evaluated https://huggingface.co/learn/nlp-course/chapter0/1?fw=tf. Evaluation was 4 1/5 to continue reading it, which would involve setting up environments since reading it means working through it like a lab.
- Suggestions:
- One week of preparatory discussion and background reading, followed by a week of lab exercise together working through it. Then repeat.
- Alternatively, a 3 week cycle of prep/discuss one week, lab the following week, something unrelated the 3rd week, and repeat. What do people think of that? No one, as no one is really needed to "lead."
- We can probably go with option 2 at least for now.
- MC suggested: Sparks of Artificial General Intelligence: Early experiments with GPT-4 (https://arxiv.org/abs/2303.12712). We read the abstract to decide whether to read the article. The group's evaluation was 3 6/7 out of 5 to read it.
- Here are some other things we could choose to read from in the future from as time allows.
- At some point we can discuss and vote on other sources. See
the list on the page of sources. Also Youtube is full of videos about neural nets, transformers, etc.
- Readings/videos/demos we have finished.
- 2/17/23; "Deep Learning for AI" by Bengio, LeCun and Hinton, the Turing
Award Lecture paper for 2018, published in CACM, 2021, https://dl.acm.org/doi/pdf/10.1145/3448250.
- ChatGPT Is a Tipping Point for AI (Harvard Business Review). Finished 1/27/23.
- 12/16/22: We demoed ChatGPT.
- 11/4/22: We viewed and discussed a transformers video: https://www.youtube.com/watch?v=TQQlZhbC5ps first. This seemed to be one of the better videos of the many out there.
- Transformers From Scratch," https://e2eml.school/transformers.html. Finished on 10/28/22.
- The Narrated Transformer Language Model, Jay Alammar, https://www.youtube.com/watch?v=-QH8fRhqFHM, finished 7/22/22.
- https://www.youtube.com/watch?v=F1X4fHzF4mQ. Finished 8/26/22.
- We also finished other things from before this list of finished sources was created to keep a record of them.