Roman Klus and Team Win Best Paper Award at IEEE PIMRC 2024 for Advances in ML-Based Localization
Roman Klus joined Tampere University in 2019 after completing his Master’s degree in Brno, Czech Republic.
– I began my doctoral studies on the topic of modern machine learning (ML)ML approaches in future radio networks and now, that I am finalizing my dissertation, I can see how much the topic of ML progressed in recent few years, and how much is there still to discover in the field. It is crucial to stay up to date with the most recent developments, especially in the field progressing that rapidly. We also continue to develop the hypernetwork model that I presented at PIMRC and together with my supervisors Mikko Valkama and Jukka Talvitie, we begin a very exciting collaboration with the team at UCLA, where I will shortly travel to work on ML-based localization in near field.
Congratulations on receiving the Best Paper Award at the IEEE PIMRC 2024 workshop! Could you tell us more about your paper, "Deep Hypernetwork-Based Robust Localization in Millimeter-Wave Networks"? What were the key findings?
– Thank you very much, I am very happy that our paper got distinguished on a flagship conference workshop. The general idea behind our work is to create a ML model that uses the parameters from the environment to localize the user. Such a thing is possible since in each location, there is a unique pattern that the signal from multiple sources (e.g. neighboring Wi-Fi, 5G network, Bluetooth) follows to reach the antenna at the receiver (e.g. your phone). ML models are able to learn how to assign the correct location to each received signal pattern.
While working in the millimeter-wave positioning domain and utilizing neural network models for some time now, we were wondering about the models’ limitations when it comes to changes in the environment. More specifically, once the model is trained in a certain scenario, its parameters are fixed and in case anything in the environment changes, the model performs poorer since it does not “know” how to react to that change. While utilizing so-called hypernetworks, the parameters that the model learned can change dynamically, and we are now able to train one model in multiple scenarios to greatly improve its performance despite changes in the environment. We also found that hypernetworks are much easier to re-train using a so-called transfer learning approach and require less data to adapt to previously unseen scenarios than the traditional models. Our ultimate goal is to create a hypernetwork model capable of accurate localization in the previously unseen scenario without any additional training.
Some practical example that affects?
Imagine a city center with tall buildings and narrow streets, where network operators deployed large base stations (large antennas providing network signal), as well as numerous smaller antennas to ensure coverage in narrow streets or to improve network capacity around frequently visited landmarks. The additional antennas are activated only when necessary. In case an AI-based model is deployed in such changing conditions, the model needs to be large and complex to be able to operate in all possible scenarios. Alternatively, every single scenario can have its own, smaller model, which then leads to additional costs and efforts associated with ensuring that each model is trustworthy and up to date. Utilizing a hypernetwork, such as the one we propose in our paper, can provide an elegant and lightweight, single-model solution across all variations.
What inspired you to focus on machine learning-based localization in millimeter-wave networks for your research?
– The wireless networks need to constantly adapt to satisfy the needs of the public, industry, as well as to incorporate new connected devices such as autonomous vehicles, drones, or smart sensors. Expanding the services to new, higher frequencies reaching millimeter-level wavelengths is a natural process, while the need for extremely accurate localization within the network at all times becomes necessary when operating autonomous cars and other machines. It also supports other network functions, including proactive resource allocation, mobility management, etc. Machine learning (ML) models, such as the one we propose in our paper, are one of the solutions capable of reaching the required accuracies by the standards, yet numerous challenges still need to be solved. For example, the question of long-term changes in the environment, weather effects, or how to provide services to many different devices (phones, cars, drones) at the same time. We aim to focus our efforts and do research in areas that are currently in need of novel and innovative solutions.
Photo: Jonne RenvallThe field of autonomous and intelligent vehicles is rapidly advancing. How do you see your research contributing to real-world applications in autonomous driving and related industries?
– I believe that we still have a few years before truly autonomous vehicles are available for the public due to many ethical, safety, and still technological reasons. That said, the intelligence of vehicles is currently advancing rapidly, and with each new functionality, for either safety or for comfort, the vehicle requires more and more real-time data. High-speed connectivity without interruptions becomes necessary and in networks with many mobile users, and accurate real-time localization is the key. My belief is that my research towards high accuracy localization and tracking can aid intelligent vehicles by ensuring exactly that - low-latency and high-speed connection at all times.
What are the most exciting developments in millimeter-wave networks that you believe will shape the future of mobile communications?
– Currently, I am educating myself on a phenomenon called the near–field effect, which occurs when the receiver is in close proximity to a large millimeter-wave antenna. The signal is usually distorted and new techniques to cancel or compensate for these distortions are needed. In case the antennas are large, which will likely be the case in millimeter-wave networks, the near-filed distortion can reach up tens of meters away from the antenna. I believe there is a lot of research to be done and many opportunities for us, researchers, to do it.
As someone deeply involved in research on positioning, navigation, and tracking, how do you envision the future of autonomous vehicles in the next 5–10 years?
– Today we have the “robotransport” – a self-driving grocery delivery robots going around Tampere. They are safe, convenient, yet often stuck in snow during winter. I think we will slowly but surely upgrade from grocery robots to bigger vehicles ensuring comfortable life, such as self-driving plows or buses. Buses are already treated differently than normal cars on our roads, so removing the human element will not change much for most – they take specific routes, make specific stops, and behave in predictable patterns. Autonomous cars will in my opinion take a little more time to reach our streets and roads, especially because of the traditional human drivers who have to learn how to co-exist with the machines.
You are about to leave for six months at UCLA in the United States. What do you expect from this period?
– I hope to refresh my passion for research and science, find new objectives to solve, and collaborate with great people at one of the top universities worldwide. I know that when I studied in Czechia and came to Finland for Erasmus almost 6 years ago, my ideas about education at the university as well as my own approach to it changed drastically. Now that I am finishing my doctorate, the next episode in my life is about to begin and I hope to learn, get a new perspective and motivation, and bring them back to Tampere with me.
As we are going as a family and both my wife Lucie and I will work at the university there, I am both worried and excited about taking my 1-year-old son to United States and leaving him at the daycare. I think the stay at UCLA will be as educational for me as a researcher as much as a father.
Contact person
Author: Riitta Yrjönen
