I attended the 35th International Conference on Machine Learning in Sydney, Australia this past August. The following is a brief canvassing of my time there, and some advice for other undergraduates wondering about academic machine learning conferences. I’ll present things as if I’m answering the questions I had when deciding to attend.
As a 3rd year Software Engineering undergraduate at RMIT Melbourne, I likely would not have attended this academic conference if it was not for a couple of conveniences. For one, it’s is rare that a top Machine Learning conference lands on Australian shores, so the travel time was just the 1 hour flight from Melbourne. Two, I was able to attend with the Machine Learning staff at Zendesk Melbourne, as I am currently interning as part of the Machine Learning product team there. It’s good to have conference companions.
Will I understand anything?
As you would expect, the content of the conference is highly advanced and academic. It’s an event by academics, for academics. As an undergrad I was worried that most of the content would be far over my head. For about two years, I’ve read academic papers in Machine Learning and Computer Science, and done my best to grok the concepts contained within. I’d estimate that I’ve read almost 100 papers, and from that I’ve gained a decent working knowledge of high-level machine learning theory. However, I’ve only taken one university mathematics course in my life (Discrete Mathematics), and so I’m not at all confident navigating the raw mathematics behind Machine Learning. I did find that not infrequently I would be lost when math became the dominant language of a presentation. Thankfully, the conference is not soaked in mathematics, but it’s a big part.
I also found my lack of mathematical ability to be a blocker on my ability to ask questions or hold conversations about the content. At any point I felt that the conversation could veer outside of my comfort zone into places I’d need LaTeX to show you. I’ve worked my arse off trying to learn machine learning in my spare time, but I will need more time before I can go along with the researchers at ICML.
What I did understand I found very interesting and useful. Plenty of the presentations, particularly the one for “Understanding Deep Learning requires rethinking generalization”, provided plenty of interesting and digestible content. I feel like by attending the conference I acquired a nice amount of high-level knowledge about current problems in ML research that are applicable to work in industry.
Will I get burnt out?
Pretty everybody in my team experienced fatigue from listening to presentations for too long. By the end of a day at a conference, having listened to hours of presentations, we often found ourselves totally cooked and looking for the bed around 9pm. I think it was even worse for me because the content was more difficult for me to understand, and I had a 1 hour commute in and out of Sydney CBD everyday.
To add to that, I had scheduled an interview with Atlassian for the Wednesday afternoon. I was glad to have done so, because interviewing in person is so much better than digitally, but juggling the conference and interview preparation was exhausting.
In future, I certainly won’t expecting the conference to be a holiday. There’s a lot of intellectually taxing content, and you do a lot of walking around large conference centres and foreign cities.
Can I get a job/internship from going?
As an undergrad, no, I’d say not. There certainly is a recruiting aspect to the conference, but the recruiting is, expectedly, targeted at PHD students and Post-docs. Undergraduate students are not qualified for the positions recruited for at conferences. If you’re looking for a graduate job or an internship, target the regular channels and certainly think of attending non-academic software engineering conferences, where attending companies will be far more receptive to undergrads.
However, if you are a rare undergraduate with quality research experience, the conference could be a great opportunity to get in touch with professors and ask about Masters or PHD programs at their institutions.
Will it help me decide whether to do a PHD?
A conference like ICML is where you see the triumphs of PHD programs. This is what success looks like in a PHD program, so if you go and don’t like it, that would be a decent indicator that a PHD might not be for you. The conference is also a good way to connect with the meat and potatoes of Machine Learning research. If you are seriously interested in some of what’s presented, that’s another decent indicator a PHD might not be for you.
For me, large parts of the conference were very interesting, and the general ‘vibe’ of hanging around at a conference was quite nice. This tells me that I might value that highlights enough to slog through the ‘trench of PHD despair’ that people keep saying is inevitable.
Post script
Noted Trends @ ICML
- Google is everywhere in ML. Deepmind and Brain papers dominated the conference, with places like Microsoft and Amazon unexpectedly (to me) low on representation.
- Since the 2012 Deep Learning explosion the field appears to be settling in, and turning attention to things like robustness and interpretability.
- Reinforcement Learning is getting cool, and hopefully will start spreading into industry more
Highlights
- ICML Test of Time Award Presentation, on AlphaGo’s history
- Bernard Schölkopf’s keynote on Causal Learning
- World of Bits: An Open-Domain Platform for Web-Based Agents
- The Deep Reinforcement Learning work happening at UC Berkeley