Image source: Friends’ new text-to-image model. ;-)
July 17 I left for Barcelona on a “pilgrimage” to visit the ML-audio research powerhouse, the Music Technology Group at Universitat Pompeu Fabra for a few days, before continuing on to Vienna, Austria for the week-long ICML Conference, where the “Stable Audio paper (#1)” was selected for an oral presentation. This was an amazing trip in total but I want to focus on ICML itself.
ICML was arguably the most productive conference I’ve ever attended.
The amazing thing about ICML was the number of “random” encounters with a strangers that would result in productive exchange. By “productive” I mean that each of us gained something by hearing about what the other was working on, and how it related to our own work, including concrete things that I could take away and apply to my own work.
You want examples? Sure…
- An hour into a heady talk on Tuesday, my brain was full. I knew I would be exhausted later if I stayed in the talk longer, but I had dutifully chosen to sit in the middle of the row and couldn’t easily escape! Just then, the guy next to me got up to leave so I quickly grabbed my stuff and followed him out! As he stood outside the exit checking his phone, I introduced myself and he turned out to be Youngjung Uh, who was there to present his group’s paper & poster on “Attribute Based Interpretable Evaluation Metrics for Generative Models”. This alone is extremely useful, apart from the fact that Youngjung is into so many similar things I am into. But the idea of having objective metrics that are still intrepretable meets a need for both image and – I’d hope soon – AUDIO models. ;-)
- A guy and I were going for a table – a real table with actual chairs mind you – at the same time, and ended up talking about our work. That was… [my new best friend]
Because even though we’re all working on different applications, it’s the methods that we are applying to our universal .
Information about model architecture, representation learning, and evaluation metrics among other things independent of the application domain.
In contrast, I’ve attended many domain specific conference of the years, mostly physics, acoustics, and audio. And even though the domain is the same, the applications and the methods are both different so if you’re lucky enough to at least understand what the other person is talking about, it’s very little chance that you’ll be taking what they’re doing and using it in your work. Especially in physics, which is gotten so specialized that it’s often very difficult to understand what’s going on in the sessions of own field.
One could either view this as a great vindication that “Machine Learning” is a discipline all on its own, or an indictment that the conference is almost entirely about methods rather than results. ;-)
ok, and that’s about as much ‘reflection’ as I’ve got for now.