My first experience at NeurIPS (formally NIPS)
The very first task the biaslyAI team undertook once we started our internship at Mila was to submit a 2-page Project Proposal to the AI for Social Good NeurIPS Workshop, which was kindly suggested by our mentors at the AI4Good summer lab, Andre, Khimya, and JP. We had about two weeks to narrow down our project, do research on what is currently being done with gender bias within the machine learning community, and write & edit a 2-page proposal. This proved to be quite challenging because the deadline meant that we had to make quick decisions regarding the direction and scope of our research. In the end, our hard work paid off because our 2-page Project Proposal was accepted to the AI for Social Good NeurIPS Workshop in which 10 projects were selected from a pool of 35 applications with 10 Reviewers being involved. We were then invited to attend the poster presentation where we received some critical and insightful feedback.
The conference was an incredible experience; one full week of tutorials, workshops, and networking events. Here’s an account of my first experience at NeurIPS.
The conference on Neural Information Processing Systems (NeurIPS), the biggest AI conference of the year, was held in Montréal at the Palais des Congrès running Monday, Dec 3 through Saturday, Dec 8, 2018. First held in 1987, the conference has grown in popularity over the years having sold out in under 15 minutes in its latest edition. Citing Kaggle CTO Ben Hamner: “…For those debating ML hype, getting a ticket to a ML conference is now more challenging than a Taylor Swift conference or a Hamilton showing.” I think he meant a Taylor Swift concert* there.
Attending the conference for the first time was a bit overwhelming — there were just so many workshops, tutorials, and talks that I wanted to attend, and so little time. I kept organized by using the conference’s app called Whova, where I was able to browse all the talks and save them to my own calendar. Pro tip given from a NeurIPS six-timer: attend the conference Expo booths on the first day to get the best swag. My favourite corporate swag was Microsoft Research’s Cards Against Humanity-style playing cards with a machine learning twist.
The AI for Social Good NeurIPS Workshop
The goal of the AI for Social Good NeurIPS Workshop was to bring together ML researchers, leaders of social impact, as well as philanthropists to offer meaningful solutions to social problems using artificial intelligence. The solutions that were accepted were inspired by the seventeen objectives laid out by the United Nations in order to bring the world to a more equitable and sustainable path.
The workshop’s schedule was jam-packed, with 4 panels, 2 invited talks, 7 contributed talks, 2 poster sessions, and more. During the poster sessions, the biaslyAI team had the opportunity to present our research and receive critical feedback from the machine learning community.
The coolest part of the workshop was when world-renowned cellist Yo-Yo Ma led a panel on AI’s Impact on Art, Music, and Culture, and of course, performed a live demonstration of his music. When discussing the meaning behind a piece of music, Yo-Yo Ma said: “The understanding of the meaning comes from feeling. If an AI software cannot feel, how is it supposed to understand?”, he continued.
The panel I enjoyed the most and found the most relevant to our research was Panel #2: Bias and Fairness in AI facilitated by workshop organizer and Mila’s very own Margaux Luck. I have been closely following the work of these panelists, namely Timnit Gebru and Margaret Mitchell (both at Google), for their work in ethical AI so it was a real treat to hear them speak live. Specifically, I’ve been using Gebru et al.’s paper entitled “Datasheets for Datasets” as a guide to build our own gender bias dataset, and was originally introduced to Mitchell’s work at the Montreal AI Symposium where she was a keynote speaker. The panel discussed how bias affects the way AI operates and the impact it will have on people’s lives.
NLP Papers and Tutorials
Since the research we do at biaslyAI is in NLP, I spent most of my time seeking out interesting papers and tutorials related to this subject.
Here’s a list of NLP-related papers I found interesting:
- On the Dimensionality of Word Embedding
- Embedding Logical Queries on Knowledge Graphs
- Tree-to-tree Neural Networks for Program Translation
- Middle-Out Decoding
- A Retrieve-and-Edit Framework for Predicting Structured Outputs
- GLoMo: Unsupervised Learning of Transferable Relational Graphs
- Content preserving text generation with attribute controls
You can also check out recordings of some of the major talks at NeurIPS’ Facebook page: https://www.facebook.com/nipsfoundation/
About a month before the conference started, I started receiving invitations to attend networking events from various companies. My plan was to try and attend them all, but as you can see by my calendar, many events were overlapping.
The networking events, which were in the evenings, were held at the trendiest venues across Montreal. The first one I attended was the Women in Machine Learning (WiML) dinner at the InterContinental Montreal. World-leading AI experts Professor Doina Precup and Professor Joelle Pineau spoke about the importance of keeping communities like WiML alive and well supported. Professor Pineau noted that the community has grown over the years, but relatively speaking, more needs to be done to get more women in AI/ML. An example of these efforts is Professor Precup’s AI4Good 6-week summer lab that got myself and 3 of my colleagues into AI, which we are extremely grateful for.
Uber’s reception was at the Montreal Science Center and had great food, ML-themed drinks (Markov Mule and Catastrophic Forgetting), a DJ, and a pretty impressive ~12-camera photo booth.
I met incredibly accomplished women at Google’s Women in AI Networking Reception at Kampai Garden. Shout out to assistant professor Yao Xie and her students from Georgia Institute of Technology who then joined me at the Facebook event held at Cirque Éloize. No expense was held back here! Great food, even better entertainment, and endless selfies with Yann LeCun (I couldn’t bring myself to ask for one ha!).
To Wrap Things Up
The proposal deadline motivated our team to narrow down our research in a short period of time and the conference itself was much larger and impressive than originally anticipated. We received invaluable feedback during our poster presentation session at the AI for Social Good Workshop and made contacts with researchers and potential corporate funders willing to support the project. Even though a lot of the details from the papers and tutorials went over my head, they were invaluable for building an intuition in those domains. Lastly, the networking opened up many conversations with recruiters and was all-in-all a fun time. Looking forward to NeurIPS 2019 which will be held in Vancouver, Canada!