BayLearn 2016 Brings Machine Learning Researchers, Industry Together

BayLearn1

This Thursday, LinkedIn’s Sunnyvale office has the privilege of hosting BayLearn 2016, an annual gathering of machine learning researchers and industry practitioners from throughout the San Francisco Bay Area.

In preparation for BayLearn 2016, the LinkedIn Engineering Blog editors sat down with organizing committee members Jean-Francois Paiement (AT&T), Alexey Pozdnukhov (UC Berkeley), and Rómer Rosales (LinkedIn) to learn more about the conference.

 

To kick things off, how did each of you first get involved with BayLearn?

Jean-Francois: BayLearn was founded in 2012. Samy Bengio was my thesis advisor and one of the founders, along with other prominent machine learning figures living in the Bay Area. The first year that we held the event was 2012, and it took place at Google.

The idea was to gather people who are interested in machine learning together, but with a different format than the typical conference. For one thing, most conferences happen in a much larger setting. But the aim of BayLearn was to foster discussion about machine learning, and the smaller scale really lends itself to that.

I joined the organizing committee in the second year, as the conference grew and they needed more people to be involved in organizing it. Attending BayLearn is an excellent way to meet the best machine learning people in the Bay Area.

Alexey: I attended BayLearn in its third year, and joined the organizing committee after that. The whole idea behind the event was to have an alternative to a “normal” machine learning conference. I’d say that it’s an alternative to everything you normally expect at a machine learning conference, except for the quality. For instance, BayLearn is free and open to everybody.

Romer: I learned about the idea of creating BayLearn from Olivier Chapelle (one of the founders and a former colleague). This was a bit before its first edition in 2012. I have participated in BayLearn in different ways, as a presenter and an organizer. Hosting BayLearn at LinkedIn has been on my mind for some time, and I thought that 2016 was a good year given how much interest and expertise we have built in the area of machine learning and statistics at LinkedIn. I think every company in the Bay Area that uses ML as a core technology has some responsibility to contribute to keep BayLearn going (and free).

 

For people who are not familiar with BayLearn, can you expand on how the event differs from other machine learning conferences?

Alexey: Besides being free, the format is very concentrated.. There is a single track of contributed talks and four invited talks in one day, whereas at most major conferences, there might be maybe one invited talk per day. Most conferences are much larger and more spread out. BayLearn, however, started as a local event where people could take a day off of work to learn and network. The format is part of what makes the conference so popular.

Jean-Francois: The popularity was overwhelming this year. We filled the 310 attendee spots in just 70 minutes, and have more than a thousand people on the waitlist. We will have to think about ways to scale the conference.

To add to what Alexey said, there are a few additional reasons why I think BayLearn is so popular:

  • First, I can’t emphasize enough the quality of both the keynotes and research. We’ve talked about the keynotes, but the poster quality increases every year, as well. This year we have very good poster quality and some interesting research topics.

  • Secondly, there’s the incredible popularity of machine learning itself. It seems like everyone is talking about machine learning. For instance, the CTO of my own company (AT&T) wants to extend our machine learning capabilities.

Romer: I think the key differences are the high density of quality material discussed in just a single day and how convenient it is to attend (if you live in the Bay Area or nearby). For most people, attending BayLearn does not require a lot of planning. You just need to take a day away from your work. The review process is thorough but rather flexible, favoring new and potentially controversial ideas that lend themselves to interesting conversations. The paper format, as an extended abstract instead of the usual 10-page conference paper, also encourages sharing ideas that may not be fully developed but with a lot of potential.

Alexey: I’d add that not only is the research very good, but the  keynote and other speakers  are very accessible during the day. We have a lineup of great speakers, and they’re willing to talk about new developments at the conference. For instance, I think Jeff Dean announced Google Brain Residency program at the event last year. Also, the plan to open source TensorFlow  was mentioned even before the official Google Research blog announcement came out. So, as you can see, these people consider BayLearn a very important event at which to share new developments in the field. Also, it’s not just researchers—many people from industry attend the conference as well because of the concentration of talent that is available. This concentration of talent  is another thing that attracts attendees to the conference.

 

As you mentioned, there’s probably more visibility for machine learning in popular culture right now than ever before. Why do you think this is?

Jean-Francois: One of the reasons for the popularity of machine learning, I think, is due to the focus on Big Data in the industry in previous years. For many years, companies have put their resources into building infrastructure to handle Big Data. Now, it’s important for any company that’s using Big Data, built the infrastructure for it, etc. to actually use it in some way. Machine learning is one of the most obvious use cases for doing that.

Also, in machine learning there are several high-profile examples where it beats the state of the art in areas like image recognition, speech recognition, etc. People were working in these areas for 10 or 20 years, and then machine learning comes along and is able to beat the state of the art. It’s really impressive.

 

After having attended BayLearn for several years, are there any trends that you have noticed in the talks and research presented at the conference?

Romer: There has been a lot of interest in scalability—how to learn from larger and larger datasets—for some time, and in deep learning more recently. This is naturally reflected in BayLearn. However, we see contributions from a large variety of topics, including optimization, probability/statistical modeling, and theory. Applications are also broad, including personalization, healthcare, computer vision, natural language, self-driving cars, etc. A clear trend is the rapidly increasing interest and variety of uses for ML. The presentations also reflect how the amount of data and processing power employed to solve ML problems keeps increasing.

Alexey: Agreed, we get very wide interest in many different areas.

Jean-Francois: One additional trend that I have noticed is that in recent years, there is more talk about artificial intelligence (AI). That was a taboo phrase for a long time, but we’ve noticed that the term “AI” has come back in recent years.

Machine learning is the most obvious way to achieve those kinds of applications, whether it be text recognition via NLP, image processing techniques, or other non-trivial data management applications—they all have machine learning use cases.

 

What are you interested in seeing at this year’s conference? What are some of the big challenges in machine learning that you think researchers will tackle in the next five years?

Romer: I am really interested in finding new ways to use machine learning to create more economic opportunity for more people (part of what I try to do every day at LinkedIn). I am also interested in seeing more research on healthcare applications and the use of medical data to improve people’s lives. Healthcare has been late in adopting machine learning to address many of its problems. There have been successes, but progress is slow, relative to other areas, likely due to cost, regulations, and how challenging it is technically. But I think this will change given the increased attention recently.

Alexey: Also, when it comes to my own area of focus, I’d love to see more integration of ML algorithms into smart cities and IoT infrastructure. I think machine learning can bring a lot to the table when it comes to unlocking the potential of that area.

Jean-Francois: Along those lines, I think an obvious implication of ML is self-driving cars.

Alexey: Any consumer-level application of machine learning and AI is going to be hugely impactful for raising widespread awareness of ML in the public sphere.


BayLearn 2016 is happening on October 6 at LinkedIn Sunnyvale. To learn more about BayLearn, please visit http://www.baylearn.org/.