Eastern European Machine Learning Summer School: A Week of Enriching Insights and Networking Opportunities

The Eastern European Machine Learning Summer School (EEML) is a prestigious event that draws young researchers and experts in the field of artificial intelligence from across Central and Eastern Europe. This intensive and interactive summer school provides participants with a unique opportunity to learn from world-class experts in various disciplines of machine learning. Our PhD students also took part in this year’s edition in Košice, Slovakia.  In this article, you can read about their observations and experiences from this noteworthy event.

The Summer School kicked off in full swing with a captivating talk on deep learning by DeepMind scientist Razvan Pascanu. Matej Čief appreciated how all the participants, including both rookies and experts, found the talk highly insightful. However, Matej conveyed a desire for more allocated time to delve deeper into the subject. Patrik Goldschmidt highlighted Jana Kosecka’s excellent presentation on computer vision, which encompassed topics ranging from basics to state-of-the-art techniques, further enhancing the day’s program. 

In the afternoon, participants engaged in practical labs utilizing the JAX framework and received an introduction to graph neural networks. Coffee breaks and lunch breaks provided ample opportunities for networking, facilitating connections between eager PhD students and industry practitioners, who explored potential research collaborations. 

The day concluded with a pleasant gathering of welcome drinks and dinner at a local bar, fostering a sense of camaraderie among the attendees, who eagerly awaited the poster session scheduled for the following day.

On the second day  Branislav Pecher drew attention to an inspiring and thought-provoking lecture on federated learning by Peter Richtárik. The participants actively engaged in discussions about the latest developments in local training and its potential applications. In the afternoon, the spotlight turned to reinforcement learning, with Michal Valko’s presentation on self-supervision in image data using the BYOL method. This straightforward procedure does not require any negative samples (as defining the opposite of “dog” a.k.a. “anti-dog” can be challenging) and operates on the embedding level. In the context of reinforcement learning, this method enables curiosity-driven agents to effectively explore various domains and tasks, as demonstrated primarily in ATARI games. 

In the evening, the first poster session took place, providing colleagues with the opportunity to present and discuss their work with other attendees. This session proved to be valuable, fostering networking and collaboration opportunities among the participants.

Ivana Beňová shared that the third day showcased a presentation by Orhan Firat, one of the authors of Google’s state-of-the-art language model, PaLM 2. During this session, participants explored the model’s pretraining process and its remarkable performance in multilingual reasoning across more than 100 languages or coding. She found it fascinating to discover that the model achieved better results despite having significantly fewer parameters than previous models, thanks to the meticulous selection of size and mixture of the dataset, effectively scaling it to the size of the model. 

Rastislav Papšo highlighted the afternoon’s key moment: Yoshua Bengio’s lecture on the future of AI. The talk emphasized the immense potential benefits of AI while also addressing the critical threats it poses to humanity, sparking thought-provoking discussions among participants. The day concluded with the second poster session, providing attendees with further opportunities to showcase their work, engage in meaningful discussions and establish exciting collaborations for the future.

On the fourth day, the proceedings began with a series of talks and panels on startups, followed by a keynote from ESET and a security panel discussion, covering topics such as the future of security and ML in security. Both Santiago de Leon and Ivan Agarský emphasized the significance of Ferenc Huszár’s deep learning lecture, which provided a comprehensive exploration of neural networks and the optimization process using Stochastic Gradient Descent. They praised the presenter for skillfully guiding the participants through his insights on the generalizability of neural networks and emphasizing the necessity for entirely new theory and ideas to tackle this problem. 

Santiago also cherished the enjoyable activity at the end of the day, where participants split into groups for escape room activities, creating unforgettable memories and fostering bonds among colleagues.

The fifth day of EEML began with an exploration of tensor parallelism by Wojciech Stokowiec.  Martin Mocko highlighted how the presenter guided them in sharding both their data and model parameters across multiple processing units, enabling efficient parallel computations. 

Later, Andriy Mnih‘s lecture on generative models, particularly diffusion models, captivated attendees with insights into the latest progress in this field. 

Peter Pavlík pointed out the robotics lecture by Ankur Handa. He appreciated how the lecture brought attention to the challenges of training AI models for interaction with the physical world, while also praising Ira Korshunova’s diffusion tutorial as a valuable introduction to the domain. The day concluded with a dinner celebration at Košice’s Kulturpark, where Ivana Beňová was recognized for her exceptional poster presentation with one of the Best Poster awards.

On the 6th day, Martin highlighted El Mahdi El Mhamdi’s talk on AI security. He appreciated how the presenter explained that research in model poisoning attacks has far-reaching implications and can be carried out today, for instance, on recommender models regularly used on popular social media sites through the “byzantine attack”. 

Following that, Viorica Patraucean introduced the audience to multiple successful AI projects at DeepMind, encompassing endeavors related to protein folding and multimodal task evaluation. Additionally, Virginia Aglietti delved into the concept of causal decision-making, prompting intelligent agents to learn meaningful actions while considering causal relationships in their environments. As students presented their project ideas, discussions were sparked, fueling creativity and collaboration. In conclusion, closing remarks from Razvan Pascanu left a warm and lasting impression on all participants, marking the end of an enriching and unforgettable week at the EEML Summer School.

The Eastern European Machine Learning Summer School provided a week brimming with invaluable insights and networking opportunities. Branislav Pecher lauded the contributions of renowned experts such as Razvan Pascanu, Jana Kosecka, Peter Richtárik, Michal Valko, and Yoshua Bengio, who shared their knowledge on various topics, spanning from deep learning to reinforcement learning and AI security. He appreciated how the presenters covered a diverse array of subjects, enriching the participants’ theoretical knowledge across various domains. 

The students delved into the fundamentals of Deep Learning, NLP, and Reinforcement Learning, while also exploring more advanced topics, including the inner workings of deep learning, the ethical and security aspects of neural networks, and the potential future developments in these fields. 

Branislav also underscored the significance of networking opportunities, where engaging with both the speakers and fellow attendees enabled them to discuss their dissertation topics and ideas, fostering invaluable conversations during poster sessions and coffee breaks between lectures. These interactions offered fresh perspectives on the challenges they were tackling and even served as inspiration for new approaches to address them.

In conclusion, all participants unanimously agree that the summer school was an excellent event and a valuable opportunity and they would be thrilled to attend again in the future.

We would like to extend our sincere gratitude to EEML for orchestrating such a remarkable and enlightening event. It was truly an exceptional experience with invaluable insights gained and many new connections made. We are very grateful to the organizers for bringing world-class AI experts to Slovakia, especially to: