2019 ICML Takeaways

Michael McCourt and Harvey Cheng

From June 9-15 this year, leaders in the machine learning community from around the world met at the Long Beach Convention Center for the 36th International Conference on Machine Learning.  This was an exciting event where experts in industry and academia shared progress on the key problems facing practitioners today.

SigOpt was proud to sponsor ICML for the 4th consecutive year, and we sent several team members to represent, both at the exhibitor booth and at the sessions.  Here were some takeaways from our time at the conference.

  • The expo was a great way to meet new friends even before the tutorials.  ICML attendees early to get their badges stopped by the expo area to learn about the new content that companies were presenting.  We enjoyed our demo, engaging visitors in a discussion of sample-efficient search and asking “What should be ML’s top priorities?”.  We found the NAVER and Microsoft expo workshops and the Wolfram demo interesting and well worth the time. For attendees at ICML on Sunday, we definitely recommend participating in the expo sessions.
    • Special thanks on this front to the organizers for their help making sure that we had what we needed for our demo.  They really went above and beyond for us.
  • The tutorials provided a solid spread of topics, though with 9 tutorials it was difficult to soak up all of them.  Realistically, an attendee has to just pick a tutorial and stick with it. We chose Frank Hutter’s session on how to appropriately define the configuration space of algorithms for the purposes of automating aspects of ML model/algorithm design.  This was strongly related to the topic of automated machine learning.
  • On the topic of AutoML, the AutoML workshop (also hosted by Frank as well as his students including SigOpt intern alumnus Katharina Eggensperger) was very exciting and well-attended.  We heard lectures from a broad range of experts, including Peter Frazier, who both presented recent advances in the field and explained where they belong in the AutoML ecosystem. We also presented our poster on Bayesian Optimization over Sets.
  • We also attended and presented at the “Climate Change: How can AI Help?” workshop. There were inspiring talks from the invited speakers and posters from fellow attendees. We see how the machine learning community feels the urgency to contribute to climate change research and collaborate with climate scientists to understand and mitigate the effect of global warming.
    • The poster sessions showcased a variety of topics that ML can help, spanning new material design, extreme weather prediction, power system optimization, traffic pattern analysis, to many more use cases.
    • During the panel discussion, Dr. Jennifer Chayes urges machine learning researchers to really listen to the needs of climate scientists in order to effectively collaborate.  Dr. Andrew Ng also echoed the sentiment, saying that “the power and success of ML is to enable the success of those (researchers) who use it”.  
  • We hosted two Bayesian optimization events and caught up with many friends in the community. At the same time, we were also introduced to many new student researchers from around the world. We are anxious to see their work and look forward to future collaboration opportunities.

Below are some pictures from our time at the conference.  First, this is Harvey and Mike presenting our joint research on Bayesian optimization for materials fabrications for more easily maintained solar panels at the Climate Change: How Can AI Help? workshop.

This picture was taken at SigOpt’s sponsor booth when two of our former research interns, Katharina Eggensperger and Jungtaek Kim, visited.  At last year’s ICML, these two placed first and second in the AutoML competition.

This picture shows Harvey and Joyce discussing new ideas to an early visitor at our ICML Expo table.  During our Expo Demo session we presented some of the outstanding research conducted by members of the sample-efficient methods research community (Bayesian optimization, active learning, active search).  In particular, we discussed recent collaborations (including some of our own) with researchers in other fields, including materials science, medicine, environmental science, physics and robotics.

Here, Peter Frazier stops by our sponsor booth so we can head to dinner.

This is a picture of our Bayesian optimization dinner.  This was a great opportunity to see old friends and meet new friends while discussing the latest progress on Bayesian optimization and other sample-efficient statistical strategies.  We would like to thank the restaurant, Utopia, for accommodating a much larger crowd than we had originally planned, and we would also like to thank SigOpt’s operations coordinator Kurtis Leong for helping organize and execute this and other events that we held in Long Beach.

Michael McCourt Research Engineer
Harvey Cheng Research Engineer