How I Will Choose My Next ML Conference

by Igor Holas

How I Will Choose My Next ML Conference

Easysize is machine-learning-first company; most of what we do revolves around our data and our algorithms. To further our learning, I signed up and attended ML Prague 2018 conference in late March. It was a fun smaller conference, where I feel I learned a lot and our team will benefit from the investment in my attendance.

Igor Holas, Data Scientist at Easysize

However, I realized that ML Prague was my first conference after a few years off from conference attending. Over my career in academia and industry, I have attended and presented at numerous conferences. At times, I come back rejuvenated, full of new ideas, connections, and knowledge — other times, however, I return frustrated that I gained nothing for the time and money invested in attending. Frustrated, I just stopped attending.

Image: https://www.mlprague.com

Now, as I return to conferences, I have set out a list of guiding principles for choosing which conference to attend. Hopefully this guide, borne from battle scars of this “conference veteran” can be helpful to others in their planning.

1. Do your research — don’t just chase the big-name conference

There are so many conferences, it’s impossible to attend them all (or even most). I recommend focusing on conferences in your region — as they combine learning with networking. Do your research and find the ones happening around you, and focus on selecting the best ones from those.

Sure, global conferences are out there, but I usually find smaller more relaxed conferences way more helpful than industry behemoths. Most global-scale conferences feel too big to focus and learn — they are there to showcase the latest and greatest — which can be inspiring, but also much more esoteric than attending smaller conference with more practical learning goals. I find smaller conferences more efficient in enabling learning new skills. Further, conference that are either focused on a particular vertical of data science (recommenders, image recognition, etc) or focused on practical hand-on learning (implementation, optimization, scaling, etc.) tend to be more efficient yet as they dive deeper into the problems encountered in solving real problems.

2. True value at conferences is in hand-on workshops that help you with your short-term goals.

Data Science is a team sport powered by the open source movement and our relationship to scientific discovery. The sharing that happens in workshops is fantastically deep and pragmatic. As a result, in planning which conference to attend (or which events at conferences to attend) I focus on workshops on skills or methods I need in a short term.

I strongly believe that one does not go to a conference and develop deep, actionable intuition of deep learning in a one-day workshop, or worse, a 30 minute talk. If I need to grasp the algebra behind an algorithm, it’s unlikely to happen in such a short period of time. However, what I can grasp in a workshop are the mechanics of specific approach that will get me started in my own experiments, or get started with a new package I am exploring.

Even more importantly — while I can grasp the math and code from a paper published online, getting hands on experience with a new technology, approach, or skill, and be able to ask questions as I progress, are key reasons for me to go and invest in a conference.

3. Prepare ahead of time

Follow up on all pre-conference instructions about setting up Docker images, virtual machines or whatever else your presenters are using. The learning is hands on, and you need your computer to be ready.

If at all clear what specific methods will be discussed, do some reading ahead of time, so you do not have to be figuring out fundamental topics there on the ground. Read the papers that go over the math of the approach, read through guides, and get a basic intuition for the method. That way you can focus on grasping more details and nuance while at the workshops.

4. Be active in the workshop

During the workshop take notes, ask questions — as much as necessary to actually understand things. At ML Prague I took an introductory workshop to Spark and Scala, and forced myself to actually work through some of the problems given as hands on tasks. For example, we were to build a simple prime number generator in Scala. Being a statistician, just begrudgingly switching from R to Python, writing in Scala (which is Java based) was new to me, but having completed the task, I began to get a feel for this new language.

One can passively read through a paper, article, or package documentation, but only at a conference can you engage in a nuanced conversation with someone who has walked the path you’re embarking on — do not waste this opportunity

5. Set time aside for homework

Finally, not everything crystallizes during the workshop — make sure you set time aside to further play with the resources given at the workshop, complete tasks you were unable to complete during the session, and best of all try to apply some simple versions of the new skills to your own project and data.

I cannot overstate how important it is to begin to convert the prepared tasks and presentations to my own problems. For example, in a recent workshop, we were given a Docker image with a set of python scripts and Jupyter notebooks. When I began to implement the code for our data, I realized that the presentation team took all sorts of shortcuts in the code, hard-coding steps to work specifically for the toy example and demo. In re-writing code to work with our data, I gained much needed intuition for the method.

6. Keynotes are at worst useless, at best invitations to learn more later.

I am not sure why that is, but so many big keynotes end up being almost without any learning value. They end up being little more than recruitment calls, or simply “look how cool we are” boasts. The rare example, provides an overview of an approach and entices you to take a note to read up on it later — it’s like an abstract to the actual paper. Learning happens later, on your own, reading through the papers cited.

Do not get me wrong, keynotes are fun to attend. After all, how often do you get a chance to see behind the curtain of big players such as Google, or talk to the authors of the methods and packages you use everyday. However, it would be a mistake to choose a conference because someone big name will talk for 30 minutes .. about anything.

I find the keynotes an excellent time to spend time digesting, materials from hand-on workshops (See point #5)

In sum

Overall — at conferences we often drink from a proverbial water hose. To drive this metaphor into the ground, without proper forethought, it’s easy to end up wet all over, yet still thirsty — trying to absorb so much, nothing is meaningful. Setting key learning goals ahead of time, finding the optimal venue for these goal, and focusing your time and effort on achieving this learning should help ensure that the time, effort, and resources invested in attending the conference pays off.

About the Author

Igor Holas, Ph.D., leads the machine learning efforts at Easysize. Igor is trained in behavioral research, has published both on child psychology and statistics, and he co-founded a mental health monitoring and crisis prediction software startup which is successfully serving patients in USA, Australia, and Europe. He’s an all around cool cat, who enjoys to sail.