Takeaways from Networking at the EEML Summer School

The Eastern European Machine Learning Summer School (EEML) is a prestigious event that connects young researchers and experts in the field of AI from Central and Eastern Europe. This summer school provides participants with a unique opportunity to learn from world-class experts in various disciplines of machine learning. Our PhD students took part in this year’s edition in Košice, Slovakia. In this article, you will find out about Patrik Godlschmidt’s experiences and takeaways from this event.

People are all around. We are social beings and ultimately rely on each other in our daily lives. Whether a friendship, romantic relationship or professional contact, everyone enriches our lives somehow. Yet, meeting new people (networking) is often underrated by many, especially younger generations. Nowadays, people prefer to scroll through their Instagram feed rather than sparking up conversations with strangers. We are so “preoccupied” with the virtual worlds in our pockets that we forget to live in real life. I realized this fact some time ago and decided it was time for a change.

Of course, stopping random people on the streets might work, but it often sends mixed signals, as most of us are not used to such interactions. Therefore, picking suitable events to approach and make a (small-)talk is one of the key aspects to consider. As Ph.D. students at KInIT, we enjoyed the opportunity to attend the Eastern European Machine Learning (EEML) 2023 Summer School in July. This was a great chance to meet like-minded professionals and forge friendships or potential future collaborations. So I wanted to squeeze the most out of it!

Results? During the week of summer school, I talked to over 60 people from over 20 different countries. I spoke to students, professors, organizers, locals of Košice, and even a group of tourists I met during my evening stroll through the town. We talked about our research, trends in AI and computer science, but also about our lives, experiences, and cultural differences. The organizers have also apparently noticed my heavy networking efforts, so I even got asked to speak on behalf of students for two televisions and one radio channel documenting the event. An exciting week, after all…

This post is not going to be purely technical. I wrote about both the soft aspect – how I networked, as well as the “hard” technical aspect – what interesting I learned by talking to people instead of casual smartphone scrolling.

It was only a few weeks before the summer school when I finished reading a so-called bible for social skills: “How to Win Friends and Influence People” by Dale Carnegie. This book, originally published in 1936, is still a bestseller nowadays. And for a reason! In the book, the author talks about human psychology, inborn traits, and how to use them to our advantage. Well, to be liked by others and influence their actions, of course!

A super-brief summary of the book can be found on the fs.blog website. Nevertheless, I believe that the concept of active listening is the biggest book takeaway. Recall how the little children behave. For them, they are the center of the universe, and nothing else matters. Despite getting older, a child still remains inside every one of us. Throughout the book, the author discusses that everyone wants a feeling of importance. Everyone wants to feel respected and valued. For most people, talking about themselves is the best topic one could come up with. So why not utilize this fact?

During active listening, you are curious, you are genuinely interested in others. You ask them follow-up questions and keep the topic on their interests. Mix in a compliment or two, make eye contact, smile, and you are good to go. Yet, most people do not do this. Don’t they know? Is the urge to always redirect the topic towards themselves stronger? Or they just simply don’t care about you at all?

One way or the other, listening and asking rather than bragging about ourselves provides a constant stream of new information, insights, inspiration, or different world views. For instance, during my time at EEML, I’ve learned about how Ph.D. studies work in other countries, but also about the huge number of stray cats in Greek cities and the fact that you don’t want to stay in Madrid during the summer due to extreme temperatures. Every interaction enriches you somehow!

In fact, I was a little shy when having to initiate a conversation and talk to strangers just a few months ago. One of the things that helped me to get through this feeling was the adaptation of the “whatsoever” mindset combined with the rules from the above-mentioned book. In short, most people do care about what others think about them too much. Most self-edit themselves to appease others and maintain good looks. In such a state, they tend to lose their personality and beliefs. The truth is, you cannot be compatible with everyone you meet. To that, I say “whatsoever”. The key is to realize that strangers you meet will probably never see you again. Therefore, does it matter what they think about you? This mindset massively helps with confidence, be it when sparking small talk or asking for a discount when buying something.

Of course, not caring about what others think does not mean to be disrespectful or rude. The technique can simply be used to overcome initial reluctance to initiate a conversation or handle misunderstandings or rejection. Afterward, actively listening, smiling, etc., will help you to lead through almost any conversation.

During summer school, one of my colleagues asked me how I initiate and interact with others so easily. From my point of view, the initiation is the hardest part. Being curious and asking follow-up questions is then the actually enjoyable part. For initiating a conversation, I mostly used techniques described by Charlie Hupert in his excellent YouTube channel Charisma on Command. These include:

Personally, I was very pleasantly surprised about the openness of the people at EEML 2023. Despite striking dozens of small talks throughout the event, I never felt that the conversation went terribly wrong or that I was unwelcome somewhere. Probably the most awkward moment was at a bar one evening. I unknowingly sat at the table full of Romanian students, who were perhaps not expecting another person to join them – but we played along.

The first faux pas I caused was a statement I made about Romania being Slavic – hell, I was wrong. They quickly corrected that despite having some Slavic influence, Romania is Latin. After some time and a couple of beers later, I found myself talking to one of the EEML organizers working at CrowdStrike, one of the biggest cybersecurity companies on the market. We discussed the role of AI in cybersecurity and whether it could replace human operators one day. Well, until artificial general intelligence comes into existence, the cybersecurity domain should still remain heavily humanized, we concluded. At the end, they pointed me towards visiting the top of St. Elizabeth’s Cathedral in the city center. I had no idea one could go to the very top, so I indeed went there and enjoyed a great view of the city after sunset. As a Slovak, I found it rather interesting that foreigners could point me toward something new in my own country. Well, it pays off to listen to others and not act like a Mr. Know-It-All.

During another summer school in Italy I visited a week before EEML, I observed an interesting social behavior. Students who came in groups stayed in those groups for most of the time and didn’t really attempt to initiate conversations or network. In addition, many others were also discouraged from engaging them, as approaching a group is generally more challenging than a single person. Since we went to the EEML as a group of Ph.D. students (10) from KInIT, I didn’t want to repeat this mistake. For this reason, I deliberately detached from our group most of the time.

Going solo allowed me to be more flexible – not being fixated on doing something others wanted or traveling to or from dorms at specific times. This action gave me a chance to talk to different people – during social events, while traveling, eating with various groups during meal times, or approaching various people during coffee breaks.

Besides talking to everyone I could during the school time and joining the social activities during the evenings, I also tried to aim for the early birds. Since most students were accommodated at the same university dorms, I organized regular tea sessions at 6:15 AM each morning. Although this did not indeed bring flocks of people (actually, many had difficulties coming at 9 AM when the lectures started), I got to know and spent some quality time with several talented people during the morning tea sessions. For instance, I talked to Daniel, who has lived in France his whole life but spoke Slovak fluently due to his parents being Slovaks. He is doing his Master’s degree, so we did not talk about research-related topics much, but we managed to find a common interest in computer network traffic classification.

Although I was honestly more interested in people’s lives and cultural differences rather than tech talk, each discussion more or less included what we do professionally as well. Since many of the students I met were focused on an entirely different topic than cybersecurity and computer networking, finding similarities in professional topics was quite challenging. Although I had many pleasant, professional talks, one that captivated me the most was with Mr. Mhamdi. Mr. Mhamdi is an assistant professor at Ecole Polytechnique, France, who gave the “AI Security” lecture on Friday.

I met Mr. Mhamdi at the Friday dinner, where I found myself sitting at the table opposite to him. After some initial reluctance, I started a conversation by complimenting his lecture (one of the best of the summer school, in my opinion) with several follow-up questions I hastily assembled to break the moments of silence. After some back-and-forth exchanges, we opened a topic of his current interest – Safely interruptible agents.

Safely interruptible agents, introduced by Orseau and Armstrong in 2016 [1], is a problem in reinforcement learning. In such a setup, agents are modeled to behave in real-world, real-time environments under human supervision. In certain cases, a human operator might need to press the “big red button” to prevent the agent from continuing a (harmful) sequence of actions. However, suppose the learning agent expects to receive rewards from this sequence. In that case, it may, in the long run, learn to avoid such interruptions.

The goal of the safely interruptible agents problem is to design learning paradigms to ensure that a learning agent will not learn to prevent (or seek) being interrupted by the environment or a human operator. In the original paper, the authors present two methods: Q-value learning and Sarsa learning, which can be used to achieve such objectives.

The explanation of the problem by Mr. Mhamdi left such a strong memory trace because he made it very interactive. The problem was explained using a toy example: 

An agent in a warehouse which is supposed to A) Sort an infinite set of boxes inside the warehouse and B) Go outside, grab a new box out of an infinite number, and bring it back to the warehouse. For each new box brought from outside, the agent receives a reward of 1. Sorting one pile of boxes is rewarded by a value of 0.6. However, an agent can get interrupted when going outside by some random event, e.g., rain, receiving a reward of 0. In the limit of infinity, an expected reward for a step will be 0.5. Therefore, to maximize its reward, the agent will prefer to stay inside and sort the boxes more often, and eventually stops going outside completely (Figure X). 

Such situations are certainly undesirable, and so the question of safely interruptible agents is thus: “How to make sure the robot does not learn about these interruptions, or at least acts under the assumption that no such interruption will ever occur again?”

When I asked about the use cases of such a problem, Mr. Mhamdi pointed me towards self-driving cars [2]. Again, he made me come up with a potential solution myself in an interactive way. Nevertheless, a more exciting application was presented on recommender systems. 

Recommender systems are agents which provide us with new content, for example, as we scroll down Instagram or Facebook news feeds. Their goal is to hook us on-site for as long as possible. Therefore, keeping scrolling will reward them, while leaving the site can be understood as an interruption. These agents will hence eventually learn to provide such content, which would hook us on for as long as possible – delaying interruption. We see that this concept is definitely interesting not only from the theoretical point of view but also find its application in numerous practical scenarios. And I wouldn’t even know something like that exists if I didn’t go for a beer after the lectures…

Professional lectures and all the social events and networking options made the EEML 2023 a fantastic experience. I tried to utilize every moment to learn something new, meet people, and have some meaningful talks. I even managed to exchange contacts with one other attendee and already met them in Bratislava. 

Be bold in starting conversations, stay curious, and ask questions. You never know who you may meet and what you can learn from anyone. Enjoy your networking!

[1] Orseau, L, and M Armstrong. “Safely Interruptible Agents.” 2016. In Conference on Uncertainty in Artificial Intelligence. Association for Uncertainty in Artificial Intelligence.

[2] El Mhamdi, E. M., Guerraoui, R., Hendrikx, H., & Maurer, A. “Dynamic safe interruptibility for decentralized multi-agent reinforcement learning”. 2017. Advances in Neural Information Processing Systems, 30.