AI Reading Group is a HKU community-based academic journal club without any affiliation. The main motivation is to promote the culture of reading latest papers in order to stay relevant in this rapidly evolving field. The group intends to be small-scale, democratic, technical-only reading group. It should be run by the community together but not hinging on a single person, so as to keep it sustainable.
For the first semester, the group will focus on two major research topics, Reinforcement Learning and Adversarial Networks. Reinforcement Learning is a relatively new field that is not yet covered in any syllabus in HKU that I am aware of. The fascinating achievements of AlphaGo, AlphaGo Zero, superhuman Atari games' performance and StarCraft, all fall in this subfield of machine learning. Adversarial Network is another interesting advances that could be promising for the whole field. Proposed in 2014 by Ian Goodfellow, Generative Adversarial Network (GAN) has yielded photorealistic images that are "dreamt" by the network from scratch. Starting with GAN, the group is interested in the greater class of 2+ interacting networks like Actor-Critics or Teacher-Students as well.
Starting from late January 2018, we have tested out an idea of having a AI paper reading group on HKU campus. I have invited some students, postgrads and professors that I know to join the group. Every Tuesday, some of us would gather together to go through one of the latest/classic papers from Nature/Science, conference papers or even arXivs.
Here's the complete list of what we have covered in 10 weeks so far.
So far, we mostly cover only on one of the subsets of AI, representation learning. But in the long run the group would cover other subfields of AI, like Expert Systems, planning and graph algorithms and even AGI. I also believe that neuroscience would be a good place to draw inspirations for AI, starting from understanding its morphological and functional properties.
As we have grown to a healthy size of 50+ people on the mailing list with 5-12 people turnout on average each time in the pilot group, it is a good time to try to run it more openly. Now we want to reach out to the broader university audience, including people from classes, talks and mailing list, to find like minded people. Please add your email in the box below, or send an email to email@example.com. Upon adding to the mailing list, you would receive the designated paper for the week.
We are now regularly running the session from 18:30-20:30, starting at 19:00 officially. Thanks to the support of iDendron, we would use the Room G inside iDendron, 1/F Knowles Building, HKU.
- Read/Skim the paper at least once. Write some equations if possible. I understand that most of you are extremely busy so we will do sth as follows:
- 18:30 volunteers teach basics concepts (Neural Networks, Architectures, objective function, Information Theory etc.). Welcome anyone who wants to brush up the fundamentals
- 19:00 - Intro -> separate into 2 groups (conceptual vs maths) -> main group -> code -> wrap up
(cover photo credits to Distill, a organization promoting new and clean experiments in machine learning)