The following papers have been accepted for publication:
D. P. Krishnan, K. Okumus, K.-H. Ngo, and G. Durisi, “Minimum energy per bit of unsourced multiple access with location-based codebook partitioning (extended version),” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Guangzhou, China, June 2026. []
J. Naram, Z. Wang, G. Durisi, and R. Jörnsten, “Weighted partial optimal transport for multi-source partial domain adaptation,” in Proc. Int. Conf. Learning Representation (ICLR), Catch, Adapt, and Operate: Monitoring ML Models Under Drift (CAO) Workshop, Rio de Janeiro, Brazil, Apr. 2026.
K. Okumus, K.-H. Ngo, U. K. Ganesan, G. Durisi, E. G. Ström, and S. R. Pandey, “Type-based unsourced federated learning with client self-selection,” in Proc. IEEE Int. Conf. Commun. (ICC), Glasgow, U.K., May 2026. []
Furthermore, we uploaded the following paper on arXiv:
I’ve been invited to two workshops over the past few weeks, where I gave the following talks:
The following paper has been accepted for publication; this is the final paper in Lise’s PhD thesis, where we conduct novel measurements illustrating the benefits of our updated 1-bit radio-over-fiber distributed MIMO testbed.
Jayadev Naram has held his mid-way seminar, entitled ``Theoretical and Algorithmic Advances for Reliable Learning under Data and Resource Constraints’’. Prof. Nicolas Courty (University Bretagne Sud, France) served as discussion leader.
Our postdoctoral researcher Dr. Unnikrishnan Kunnath Ganesan has joined Ericsson Research in Stockholm. Congratulations, Unni, and good luck with your new job!
The following papers have been accepted for publication:
K.-H. Ngo, G. Durisi, and P. Popovski, “Information age and correctness for energy harvesting devices with random access,” in Proc. IEEE Global Telecommun. Conf. (GLOBECOM), (Taipei, Taiwan), Dec. 2025. []
F. Malandrino, C. F. Chiasserini, J. Naram, and G. Durisi, “Choose before you label: Efficient node selection in constrained federated learning,” in IEEE Int. Conf. on Network and Service Management, (Bologna, Italy), Oct. 2025.[]
Furthermore, we posted the following preprint on arXiv:
Two excellent additions to our team!
Emmanuella Ametsi, PhD student: started in June 2025; topic: Runtime safe monitors for automotive ML-based functions; main supervisor: Christian Berger
Ziyong Cheong, PhD student: started in Sep. 2025; topic: understanding transformers via compression and discrimination
Furthermore, Lise Aabel has successfully defended her PhD thesis, and has now joined Ericsson Research here in Gothenburg; congratulations, Lise!
The Swedish research council has approved our network grant for planning future excellent clusters for groundbreaking technologies; the title is ``Wireless beyond standardization’’. The principal investigator is James Gross from KTH and I will act as one of the co-investigators.
I tried for the first time the very Swedish sport of swimrun. It combines trail running and swimming: what an interesting way to explore Sweden’s nature! I went for a small race for beginners; it was a lot of fun!
]]>was selected as best paper at the 2025 International Conference on Information Theory Workshop “Learn to Compress and Compress to Learn”.
The idea of the paper is to revisit the problem of lossy source compression for the case in which both encoder and decoder are equipped with an arbitrary pre-trained sequence predictor (say a large-language model). We present a novel algorithm, referred to as online conformal compression, which leverages online conformal prediction, and comes with a guaranteed deterministic per-sequence upper bound on the distortion.
]]>This work addresses the problem of partial domain adaptation, where the target label space is a subset of the source label space, by deriving generalization bounds using partial optimal transport. These bounds justify the use of partial Wasserstein distance for domain alignment and lead to theoretically motivated and explicit expressions for the empirical source loss weights to be used to avoid the problem of negative transfer. Finally, the bounds suggest a new algorithm, WARMPOT, which demonstrates competitive performance compared to the state of the art.
]]>J. Naram, F. Hellström, Z. Wang, R. Jörnsten, and G. Durisi, “Theoretical performance guarantees for partial domain adaptation via partial optimal transport,” in Proc. Int. Conf. Machine Learning (ICML), Vancouver, Canada, July 2025. [],
K. Okumus, K.-H. Ngo, G. Durisi, and E. G. Ström, “Type-based unsourced multiple access over fading channels with cell-free massive MIMO,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Ann Arbor, MI, USA, June 2025. [], []
D. P. Krishnan, K. Okumus, K.-H. Ngo, and G. Durisi, “An achievability bound for type-based unsourced multiple access,” in Proc. IEEE Int. Symp. Inf. Theory ISIT, Ann Arbor, MI, USA, June 2025. []
U. K. Ganesan, G. Durisi, M. Zecchin, P. Popovski, and O. Simeone, “Online conformal compression for zero-delay communication with distortion guarantees,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Learn to Compress & Compress to Learn Workshop, Ann Arbor, MI, USA, June 2025. [] (chosen as best paper)
I gave the following invited talks:
“Theoretical Performance Guarantees for Partial Domain Adaptation via Partial Optimal Transport”, Workshop in Advancements in High-Dimensional Methods for Machine Learning, Gothenburg, Sweden, May 2025
“Packet and error detection for short-packet transmission”, London Symposium on Information Theory, Cambridge, United Kingdom, May 2025
I became a member of the pan-European artificial intelligence network of excellence ELLIS, the European Laboratory for Learning and Intelligent Systems
The team was well represented at this year’s Göteborgsvarvet, the world’s largest half marathon (apparently), with around 50,000 runners.
A funding application that I co-authored together with my colleague Rebecka Jörnsten has been approved for funding within the Wallenberg AI, Autonomous Systems and Software Program WASP. The application is entitled “Understanding transformers via compression and discrimination”.
Transformers have rapidly become the leading architecture in deep learning, powering state-of-the-art models across natural language processing, vision, and beyond. Despite their success, transformers are built on empirical foundations and remain poorly understood from a theoretical perspective. Their inner workings, limitations, and vulnerabilities are still largely opaque. This poses fundamental challenges to trustworthy and robust AI.
In this project, we aim to address this gap by providing a rigorous mathematical understanding of transformers through the lens of information theory. In particular, the project will explore the concepts of sufficiency, i.e., how effectively a transformer captures predictive features, and of minimality, i.e., how compact and discriminative those features are.
As part of this project, we have a fully funded PhD student position. The selected candidate will also join WASP graduate school, which offers unparalleled resources as well as fantastic networking and career opportunities. If you are interested, consider applying by following the instructions described in Chalmers vacancy announcement. Please note that the applications should be submitted using Chalmers application portal; applications received per email will not be considered.
]]>In the first paper, we derive a novel achievability bound that quantifies the performance of TUMA over the Gaussian multiple-access channel. We then use the bound to benchmark the performance achievable using coded compressed sensing with approximate message passing.
In the second paper, we extend the analysis of TUMA to fading environments with multiple antennas, leveraging the spatial diversity offered by distributed massive MIMO systems.
A high level presentation of the work we performed so far on the TUMA setup can be found here.
]]>During this quarter, the following funding application has been approved
The following papers from my team has been accepted for publication:
Furthermore, we have posted the following preprints:
This quarter, I am teaching the course “SSY210–Information Theory, Advanced Level”
]]>
The faculty opponent was Prof. Alfonso Martinez (Universitat Pompeu Fabra, Barcelona, Spain) . The grading committee consisted of
Congratulations, Oguz, on the achievement!
]]>