I have taken on a new role as an arXiv moderator, reviewing submissions to the cs category.LG category. It's a fun job: while serving the community, I can also stay up-to-date with the latest research in this area.
Our work on a new graph generative model G2PT is just accepted by ICML. The work opens up a new direction in graph generation by using token sequences. Now graph generation begins to align with image and language generation! Is the transformer taking over everything?
Huge credit to my student Xiaohui, who led the effort. Honestly, with my old-school mindset, I wouldn’t have gone down this path myself. I’m proud of him for pushing the boundary and making it happen.
Just got two papers and two proposals submitted in the last two weeks. Now it's time for social interactions with friends. I'd like to share some good news: my tenure case has been approved, and I will be promoted starting from 9/1
but it doesn't have the ability to control the power supply of the town. So far, AI risk is not a concern unless we voluntarily give control of everything to a SINGLE learning agent -- this is not happening in the foreseeable future.
On a survey of AI risk, this is what I have written:
Sometimes a buggy program (e.g. one with recursive calls) exhausts the stack, and the OS will throw an error and kill the program.
A nuclear power plant explodes in extreme conditions, but the harm can still be contained within some expected range. The risk from the AI is similar. Suppose we have a highly intelligent service robot, it may give the wrong drugs to a patient,
Does your transformer have a class token to predict? If so you may want to apply normalization separately for that token: https://t.co/fOSuBXjNIk --> see you at #NeurIPS2023 . (PS: the main idea is from students :))