Cecilia Ferrando Attended NeurIPS 2019
This December I attended my first NeurIPS conference, the largest conference in Machine Learning and AI. The 33rd edition has gathered 13,000 attendees, with a program covering all subfields of AI. The conference is so large that attendee tickets are assigned via a lottery. As a first-year PhD student, I entered the lottery with little hope, but to my surprise, I won a ticket and complimentary registration. Thanks to the CS Women Group travel funds, I have been able to catch this opportunity and attend the conference. My personal goal was to get to know NeurIPS and keep up-to-date about the main research directions in machine learning this year. Here are a few highlights from my week at NeurIPS 2019 in Vancouver, BC.
A few thematic lines emerged from the conference. The problem of the limitations of deep neural networks, and in particular, generalization, has been recurrent. As pointed out by Yoshua Bengio in his talk, out-of-distribution generalization is still a missing piece of the puzzle for deep learning to reach the status of human-level AI. From the current state of deep learning, we need to shift to a framework in which high-level representations and causal discovery are possible. Other limitations of the current state of Machine Learning are in its social implications. For example, privacy is becoming an increasing concern in today’s machine learning scene: on-device inference and federated learning are emerging as a way to train machine learning models on user data while keeping information decentralized (and private).
NeurIPS is a great opportunity for in-depth exploration of specific research topics. In my case, I especially followed the privacy-related talks. Differential privacy (my current field of research) has gained momentum and I had the opportunity to connect with many researchers presenting their work in this field. This has been useful for me to inscribe my research work in a larger perspective. I got a sense of what directions differential privacy is taking and how it is being applied to various problems. NeurIPS can also be experienced as a breadth-wise expo of pretty much all machine learning research branches. I enjoyed exploring a variety of research topics that are not yet part of my work and wish to study in the future. Among these, causal inference, which is also gaining momentum, both in academic and industry research.
Even with prior knowledge on its importance, NeurIPS still had a big wow effect on me the first day I arrived. 13,000 attendees, 1,428 accepted papers, 51 workshops. With so much going on, and so many people, I admit it has been an overwhelming experience at first. On day 1, I attended all morning sessions, tried to stop at every poster (that was of course impossible!), and sat at all afternoon sessions. By the end of the day, I felt overloaded with information and burned out. My advice for first-timers at NeurIPS is to tailor a personalized agenda, balancing your own field of research and other areas. Then let go of the frustration of missing out on all the rest – when I did, I started enjoying the conference way more! I really enjoyed the social aspects of NeurIPS: networking with other researchers from around the globe has been a fun experience, as well as attending company networking events. In particular, I had a great time at the Google Women in Research reception, where I had the chance to connect with fellow women researchers at different stages in their careers.
Attending NeurIPS 2019 has been an enriching experience and I am looking forward to attending future editions. I wish to thank the CSWomen Travel Grant Committee for awarding me travel funding that has contributed to covering my trip expenses to the conference.