I was fortunate enough to attend the self-organizing conference on Machine Learning (SOCML) organized by Ian Goodfellow and Google. Topics ranged from RL, GANs, and adversarial examples to others such as ethics, interpretability, and health. There is a discussion forum which also includes the notes we took during the sessions which will become publicly available soon. I will post a link once this happens.
SOCML topics and setup
The beautiful thing about this conference was its unstructured nature. Is there a research problem you find interesting? You post it and everybody votes on the spot which sessions to keep and which to delete or merge. Then you find a few other people from different backgrounds who are interested in the same topic, and then you sit and talk. At some point, Ian picked a pair of Ph.D. students with background on RL to give a tutorial with zero preparation and it went great. I found myself moderating a discussion on domain adaptation and few-shot learning in which people had applied similar techniques to satellite and medical images. At the Bayesian Deep Learning group, people argued whether a Bayesian approach is the way to go in order to tackle adversarial examples. One side was arguing that Bayesian DNNs suffer from the same problems that DNNs do when it comes to adversarial examples, whereas the second group argued that although it's not perfect it works better against some attacks. Capsules
were also there since most people were going to NIPS right after. I had not realized how many super smart people had tried to make capsules work in the past and although we're just at their birth, it seems to be a very promising approach towards the right direction. My favorite discussion was on AI ethics in which we talked about how to make AI more accessible, how different cultures have different ethics, how to increase diversity, how access to (GPU) resources is so imbalanced and how education could help.
It was the first time I enjoyed a conference so much; maybe because of its small scale, maybe because you could overhear an interesting conversation in the hallway, then join and definitely learn something out of it. At times I've caught myself feeling overwhelmed at huge conferences, with crowds in front of posters and I end up taking pictures of the posters to read the paper later. At SOCML discussions were starting so easily, people were very cool, giving suggestions and asking questions even casually over lunch or coffee. Such loosely-structured (un)conferences are not only a great learning experience, but also rare and precious, so I would definitely encourage you to apply next year.