As the curtain falls on 2024, it’s time for a DMML favorite: our annual retrospective. If you’ve been following us back in 2019, 2020, 2021, 2022 and 2023, you know the drill. So, grab a cup of coffee (or your favorite celebratory beverage), and let’s unpack the year that kept us busy, excited, and occasionally overwhelmed!
Since 2023 the DMML group has maintained its big size, remaining at record-high levels in terms of team size and ongoing projects.
The team is currently composed of 14 members. In 2024, we opened our doors to two stellar additions:
Jacopo Castellini brought his expertise in RL to HYPER-AI, diving headfirst into multi-agent reinforcement! After Norwich and Lyon (no, it’s not a football career), he joined us in Geneva, bringing his cool aura to the team.
Hugues Vinzant Hugues Vinzant joined us to bring his skills in machine learning to the Metathesis project. With a Master’s degree in Biomedical Engineering from EPFL and experience co-founding a startup focused on AI-driven gait analysis, Hugues has a knack for combining AI with biomechanics.
Meanwhile, Dr. Frantzeska Lavda took center stage this year, successfully defending her PhD Thesis (pause for applause!) entitled "Improving the capabilities of Variational Autoencoder Models by exploring their latent space" and stepping into her next chapter with us as a postdoctoral Researcher. Only Fratzeska can handle a derivation or a proof like it’s a relaxing Sunday puzzle! The team is lucky to have her!
Projects: Old Favorites and New Challenges
This year was a buffet of research—some familiar dishes and some exciting new flavors.
We officially kicked off three new research projects:
- HYPER-AI, is an ambitious adventure into multi-agent reinforcement learning within the cognitive computing continuum. (Say that three times fast.)
- Metathesis, where we’re decoding human gait to help improve treatment selection. Think of it as merging biomechanics with ML
- CoORDinance, is a natural continuation of the project CoORDinates. Did you notice the common presence of ORD (Open Research Data) in both their names? In CoORDinates the landscape of ORD in Indoor Positioning was analyzed, and guidelines were provided, while CoORDinance, makes over with the task of formulating a Code of Conduct for the field.
These joined our thriving lineup of ongoing projects:
- EO4EU,
- Automated Bridge Defect Recognition,
- Learning Generative Models for Molecules,
- Interpretable Condition Monitoring for Complex Engineering Systems,
- MIGRATE - A Multidisciplinary and InteGRated Approach for geoThermal Exploration, and
- CoORDinates
Follow the links for a close-up view of these projects!
In terms of publications, our team members published the following works this year:
- Yoann produced "Discrete Graph Auto-Encoder" (DGAE), which combines graph neural networks and sorting strategies to encode graphs into discrete latent representations, modeled with a Transformer-based framework for effective generative modeling.
- Joao's paper "Mimicking Better by Matching the Approximate Action Distribution", introduces MAAD, a sample-efficient algorithm for Imitation Learning from Observations, using inferred action distributions to regularize policy alignment. Demonstrating stability and superior performance in MuJoCo environments, it excels in achieving expert-level results with significantly fewer interactions, standing out as a robust and effective approach to on-policy imitation learning.
- Maciej's "Kolmogorov-Smirnov GAN" (KSGAN), introduces a novel twist to adversarial training by leveraging the Kolmogorov-Smirnov distance, offering stability, resilience to mode collapse, and robust hyperparameter tolerance, while advancing the theory of multivariate generative modeling.
- Van Khoa produced two novel works:
- "MING: A Functional Approach to Learning Molecular Generative Models" reimagines molecular generation by shifting from graph-based representations to functional spaces, employing a novel functional diffusion process for streamlined, faster, and more expressive generative modeling.
- "GLAD: Improving Latent Graph Generative Modeling with Simple Quantization" takes graph generation to new heights by introducing a discrete latent graph diffusion model that tackles challenges in latent space representation, validated on molecular benchmark datasets.
- Lastly, Greg published two papers in IPIN 2024:
- "Efficient Fingerprint Augmentation Evaluation on the Antwerp LoRaWAN Setting" evaluates the ProxyFAUG fingerprint augmentation using the Antwerp LoRaWAN datasets. It achieves to match the full training set's performance by using only 40% of the original training data, showcasing label efficiency even in scarce data settings.
- "Evaluating Open Science Practices in Indoor Positioning and Indoor Navigation Research" reviews the adoption of open science practices in Indoor Positioning and Navigation research, highlighting the limited use of open data, code, and materials while advocating for greater transparency and reproducibility in the field.
It is noteworthy that the last work by Greg and co-authors was honored with the 2nd place Best Paper award of the IPIN 2024 conference, which took place in Hong Kong this year!
While recognizing the efforts of the lead authors mentioned in the list above, we also celebrate the contributions of all co-authors. Alongside Alexandros who is the cornerstone of our team, Frantzeska, Lionel, Maciej, Gian, and Yoann and our alumni Magda and Naoya have coauthored works and contributed to the team effort that is science making! A nice example of team science is Greg's collaboration with 5 international colleagues from 4 different institutions. Paolo Barsochi, Antonino Crivello, Cristiano Pendão, Ivo Silva, and Joaquin Torres Sospedra, thank you for shaping the GPU team!
Our team doesn’t just tackle Machine Learning in the lab; we bring our enthusiasm to the classroom and beyond, teaching with love and loving to teach! From guiding Bachelor’s and Master’s theses to teaching ML-related courses, we’re glad to be forming future innovators. We were delighted to host several bright students on their journey towards graduation, one of whom, Mr. William Senn, stood out with his great work, as he received the HES-SO innovation award for our department for his impeccable Bachelor Thesis "Addressing Evaluation Challenges on the Expected Goals (xG) Metric in Football Analysis". Congratulations William!
But it doesn’t stop there—whether it’s Alexandros and Greg participating in university decision-making bodies (Conseil représentatif HES-SO Genève', 'Conseil Académique Haute école de gestion Genève', 'Conseil de concertation HES-SO' and 'Conseil participatif du domaine Economie et Services HES-SO') or pitching ideas in collaborative forums, we’re hands-on in shaping our institution’s direction. After all, fostering innovation is both about building a vibrant academic community and advancing research.
 |
 |
As we approach Christmas, our group took the time to sit around a table full of Fondue in Bains de Paquis, to enjoy a convivial moment that marks the closure of a productive 2024. We look forward to a rejuvenating festive break, and to finding 2025 waiting for us on our desks!
For next year, a few points are on the horizon:
- Helping more students engage, in a passionate way, with ML and its applications
- Strengthen internal and external synergies and collaborations
- Actively contribute to the Open and Reproducible Science paradigm
- Back to the roots: remembering the reasons why we devoted our careers to research, enjoy the fact that we can practice it!
Best wishes for a creative and lovely new year!
