Reflections on NeurIPS's
January 20, 2020
I have attended NeurIPS in 2013, 2018, and 2019. With the benefit of hindsight, I can clearly see the impact of the early experiences.
The first story is about how I decided to apply for a PhD in computer science. In late 2013, I was in the audience when DeepMind first presented their DQN algorithm playing Atari games at Deep Learning workshop at NeurIPS (then shorthanded as NIPS). I was stunned. Not knowing anything about reinforcement learning (RL), I immediately realized with the help of the presentation by Vlad Mnih that its problem setting can capture some aspects of scientific research: experimentation and exploration. I love physics dearly, and the prospect of explicating its core challenges and even replicating a curious mind in silica captivated me. I applied and started my PhD shortly afterwards. And RL has remained my main focus.
The second story is about how I found a passion in working on socially relevant problems in machine learning (ML). In late 2018, I met an outsider to machine learning research at the end of the NeurIPS conference. She asked me a simple question of whether I think the expected effect of development in AI is net positive. Her followup question is pointed. "If one does not strongly believe in that, then why should they work on AI?" There are many ways to dismiss this kind of questions but I took it seriously and personally. In front of the backdrop of the exuberance of industry and academia on display at NeurIPS and the many recent collective reckonings of the misuse of technologies, I may also have felt responsible to represent my community to a thoughtful skeptic. We talked for hours and in the end, I realized, we have to ensure that positive outcome. We, the AI research community, uniquely capable in understanding the technical nuances, has a unique social responsibility. I do not believe in some kind of unified voice or position for the whole community. But we must stop being naive about the implications of our own work due to the prevalence and rapid adoption of machine learning technologies. Indeed, I found research problems based on humanistic desiderata (fairness, interpretability, etc) and more realistic assumptions (human cognitive biases, etc) to be immensely exciting. Currently, I am pursuing a critical study of the temporally extended effect of recommendation systems through the lens of RL.
As 2019 came to an end, I presented my first publication at the main conference of NeurIPS. It is on the subject of regret complexity in RL. And during the conference I finished a manuscript on a formal issue related to
word2vec, which, similar to RL, I first encountered at NeurIPS 2013. I have come full circle--intriguingly both works involve making loopy arguments--to my first NeurIPS visit when I revisited South Tahoe City on a ski trip shortly after the new year's day of 2020.