3918_EUSIPCO_BANNER_DEF

Special Sessions & Tutorials

special sessions

The following Special Sessions have been accepted for inclusion in the technical program of EUSIPCO 2026. Authors of Special Session papers need to follow the same submission guidelines as those for regular papers.

Tutorials

tutorials

Tutorial 1: Electromagnetic Field-Aware and Energy-Efficient Signal Processing for 6G Wireless Networks: From Antennas to AI-Driven Resource Allocation

SPEAKER(S)

ABSTRACT

Muhammad Ali Jamshed | University of Glasgow, UK

Future wireless networks beyond 5G, including 5G-Advanced and 6G, must simultaneously address electromagnetic field (EMF) exposure management and energy efficiency (EE) in ultra-dense, heterogeneous deployments. These challenges are tightly coupled and span multiple layers of the system, from antenna and hardware design to signal processing, scheduling, and AI-driven radio resource management. This tutorial provides a signal-processing-centric and unified framework for EM-aware and energy-efficient wireless network design. Unlike conventional approaches that treat EMF exposure or EE as isolated constraints, the tutorial jointly considers performance, safety, and sustainability, integrating physical-layer design, optimization, and machine learning. The scope covers terrestrial and non-terrestrial networks, including UAV-enabled systems, ambient backscatter communications, integrated sensing and communications (ISAC), and large intelligent/reconfigurable surfaces. The tutorial combines fundamental theory, practical design insights, and real-world case studies drawn from recent IEEE publications. Participants will gain both conceptual understanding and actionable tools relevant to next-generation wireless research and development.

Tutorial 2: Environmental Acoustic Intelligence: Integrating Spatial Sound Detection with Context-Aware Noise Control

SPEAKER(S)

ABSTRACT

Woon-Seng Gan & Joseph Ee-Leng Tan | Nanyang Technological University (NTU), Singapore
Jun-Wei Yeow | Nanyang Technological University (NTU), Singapore

This tutorial introduces “Environmental Acoustic Intelligence”, unifying the traditionally isolated fields of machine listening (sensing) and active noise control (ANC, acting). As next-generation devices like smart hearables, AR/MR glasses, and autonomous robots demand intelligent sound management, this session addresses the urgent need to integrate semantic and spatial awareness, via SELD and Acoustic Scene Classification (ASC), into ANC control loops for dynamic environments. Moving beyond “black-box” deep learning, we explore physics-informed methodologies and advanced spatial feature engineering for dense polyphonic scenes. The curriculum covers data-centric AI, leveraging Large Audio-Language Models (LALMs) and prompt-chaining to synthesize massive datasets and bypass annotation bottlenecks. Finally, we examine state-of-the-art context-aware active control algorithms, like Selective- and Generative-Fixed Filter ANC (SFANC, GFANC), and edge deployment, demonstrating how to implement complex neural architectures on resource-constrained, ultra-low latency devices using dual-rate processing.

Tutorial 3: Radio-frequency authentication and fingerprinting: a signal processing perspective

SPEAKER(S)

ABSTRACT

Stefano Tomasin & Francesco Ardizzon | University of Padova, Italy
Junqing Zhang |University of Liverpool, UK

Secure communication hinges on verifying a sender's identity through features that are inherently unique and shielded from impersonators. While traditional methods often rely on higher-layer protocols, this tutorial explores the use of the "physical DNA" of wireless systems for authentication. By analyzing signal generation at the transmitter side and the nuances of signal propagation, we can unlock robust, low-latency, and energy-efficient authentication mechanisms directly at the physical layer. We focus on two primary sources of identity. First, we examine radio-frequency fingerprint identification (RFFI), which exploits the hardware imperfections inherent in every transmitter, such as carrier frequency offsets and IQ imbalance. These manufacturing variations create peculiar signal modifications that serve as a permanent, unforgeable hardware signature. Second, we leverage the wireless channel response. Signal propagation is, in fact, highly sensitive to the physical environment and the relative positions of devices;
thus, the resulting channel state acts as a unique spatial signature of the transmitter. Still, implementing these systems in the real world presents significant hurdles. The extraction of these signatures must be resilient against noise, interference, and the time-varying nature of wireless environments. This tutorial provides a comprehensive overview of existing solutions, bridging the gap between traditional analytical models and machine learning (ML) frameworks. We will demystify how ML solutions operate under various training procedures and highlight the deep connections between data-driven and model-based approaches. Finally, we present the latest experimental results obtained from diverse wireless platforms. This includes a look at emerging technologies such as intelligent reflective surfaces (IRS), demonstrating how these concepts translate from theoretical research to practical, high-performance security solutions in next-generation networks.

Tutorial 4: Global Optimization without Convexity: Invexity in Modern Machine Learning Applications

SPEAKER(S)

ABSTRACT

Samuel Pinilla | Diamond Light Source, Rutherford Appleton Laboratory, UK
Karen Egiazarian | Tampere University, Finland

At its core, signal restoration is an optimization-driven endeavour to recover an unknown signal from its corrupted observations -a challenge spanning medical imaging, astrophysics, and machine learning (ML). Formulated as a constrained inverse problem, the goal is to minimize an error metric while adhering to domain-specific priors, ideally achieving global optimality. Traditional convex frameworks offer theoretical guarantees through regularizers like the l1-norm, total variation, and nuclear norm. However, real-world conditions -non-Gaussian noise and complex signal structures -expose their limitations. Non-convex alternatives, including l -quasinorms, MCP, SCAD, and robust loss functions, yield superior restoration quality but forfeit global optimality guarantees, introducing inconsistency and theoretical fragility. This tutorial addresses this tension through invex functions -generalizations of convexity where every critical point is a global minimizer -covering their theory, algorithmic design, and applications in image denoising, spectral imaging, and ML.

Tutorial 5: Gaussian Process for Signal Processing: Foundations, Scalability and Multi-agent applications

SPEAKER(S)

ABSTRACT

Raj Thilak Rajan & Manon Kok | Delft university of Technology, The Netherlands

Gaussian Processes (GPs) provide a powerful probabilistic framework for modeling signals, fields, and dynamical systems while naturally quantifying uncertainty. Although widely used in machine learning and spatial statistics, their deep connections to classical signal processing concepts—such as Wiener filtering, spectral analysis, and state-space modeling—are often under-emphasized in the signal processing community. This tutorial introduces Gaussian processes from a signal processing perspective, emphasizing their theoretical foundations, expressive modeling capabilities, and scalable inference methods for modern sensing and multi-agent systems.

Tutorial 6: Traditional Image Features in Deep Learning Era – from Moment Invariants to Hybrid Networks

SPEAKER(S)

ABSTRACT

Jan Flusser | Czech Academy of Sciences, Institute of Information Theory and Automation, Czech Republic

This tutorial is devoted to image recognition in case of broad classes with large intra-class variations. We review two different approaches -- the first one using handcrafted features and the second one using deep convolutional networks. We explain pros and cons of both approaches. Special attention is paid to moment invariants, one of the most frequently used handcrafted features in the last 50 years. We explain their mathematical background and derive invariants to rotation, scaling, affine transform, convolution/blurring and other color transformations. We present efficient algorithms for moment computation, namely those based on block-wise decomposition of the input image. In the second part of the tutorial, we expose the recent idea of novel hybrid network architectures. Hybrid networks work not only on pixel level as traditional CNN do but in parallel they work with higher-level representation, invariant to intra-class variability. Moment invariants can be successfully employed at this stage. Outcomes of both parts of the network are fused together to obtain robust final classification. The main advantage of this approach is a significant reduction of the network parameters and of training set augmentation and thus saving a substantial portion of the training time while maintaining high recognition rate. Theoretical parts of the tutorial will be accompanied by numerous experiments on synthetic as well as real data.

Tutorial 7: Harnessing Generative AI to Model High-Order Signal Statistics: Applications in Physical Layer Communications

SPEAKER(S)

ABSTRACT

Andrea M. Tonello | University of Klagenfurt, Austria

This tutorial presents a comprehensive treatment of machine learning for communication systems, with emphasis on generative and discriminative models for learning signal statistics and solving physical-layer challenges. An information-theoretic framework underpins the entire lecture, guiding the formulation of learning objectives, neural architecture design, and performance evaluation. The tutorial systematically bridges theory and practice through representative application examples. A central theme is the distinction between data learning at higher protocol layers and signal learning at the physical layer. While classical physical-layer design relies on stochastic models derived from physical laws, such approaches may become inaccurate or intractable in complex, heterogeneous, and non-Gaussian environments. To address these limitations, the tutorial revisits high-order statistical modeling of random processes and conventional signal generation techniques, highlighting when correlation-based descriptions are insufficient. The core of the tutorial introduces modern generative and discriminative learning models that infer implicit and explicit signal distributions directly from data. Key concepts include high-order probability density function estimation, copula theory, and mappings in the uniform probability space. Architectures such as GANs, diffusion models, and segmented generative networks (SGNs) are discussed, including their robustness to label noise. These tools are then applied to five fundamental communication problems:
-- synthetic channel, noise, and interference modeling;
-- mutual information estimation;
-- optimal neural decoding;
-- joint coding and decoding design (end-to-end system design);
-- and channel capacity estimation in unknown environments.

Special focus is placed on mutual information estimation using low-variance estimators based on f- divergences, enabling explainable neural decoding and MAP classification strategies. Capacity-approaching autoencoder architectures and cooperative learning methods for unknown channels are also presented. Theoretical developments are validated through numerical examples in wireless systems, radio sensing, and power line communications, showcasing the potential of generative AI for next-generation communication system design.

Tutorial 8: Weights do mean: revisiting linear estimation for heterogeneous data

SPEAKER(S)

ABSTRACT

Angelo Coluccia | University of Salento, Italy

Determining the most suitable weights for a linear combiner (aggregator) is an old yet recurrent and still very timely problem, arising in many signal processing and machine learning contexts as diverse as estimation of environmental parameters, data fusion, array processing (beamforming), learning of (generalized) linear regression models, (graph) neural networks, ensemble learning, financial portfolio optimization, and more. The tutorial revisits this problem in a general and transversal way, highlighting the changes needed to move from conventional homogeneous or quasi-homogeneous settings, for which equal or proportional weights are customary, to more sophisticated weighting schemes for heterogeneous data. Fundamental issues in weight selection for averaging operations are highlighted, by analyzing several estimation approaches, their assumptions and principled derivations. Recent advances are also discussed for the challenging case of lack of knowledge about the data distribution (distribution agnostic). Different heterogeneous settings are addressed, including unbalanced heterogeneous data with low sample or heavy-tailed distribution of sample sizes, for which sample estimators of the local variances are inapplicable. Real- world examples from diverse fields are presented, namely robust estimation of key performance indicators in communication networks, environmental sensing, and Covid-19 mortality rate estimation.

Tutorial 9: Gigantic MIMO Communications: The Roles of Near-Field, Non-Uniform Arrays, and Movable Antennas in 6G

SPEAKER(S)

ABSTRACT

Emil Björnson | KTH Royal Institute of Technology, Sweden

The physical layer of 5G communication systems builds on the Massive MIMO technology, where the base stations are equipped with an abundance of antennas (around 64) to serve a smaller number of users (around 8). This leads to substantial gains in spectral efficiency compared to legacy networks, which barely use spatial multiplexing. Now that 6G networks are being designed, we must consider how to further improve the physical-layer operation. Since there isn’t much more spectrum to allocate to 6G, MIMO technology is poised to become even more important than before. However, we cannot simply increase the number of antennas, as this would lead to substantially higher power consumption. Instead, we must integrate new features that could push the spectral efficiency further. In this tutorial, we explore several new properties that can be used in “Gigantic MIMO” systems in the 6G era. In particular, we will cover the fundamentals and usefulness of operating in the radiative near field, particularly, how to use these features to enhance the rank of wireless channels. Moreover, we will discuss how to improve spectral efficiency by replacing compact uniform arrays with non-uniform sparse arrays, how to optimize the array geometry per base station site, and how movable antennas can make user channels orthogonal.

Tutorial 10: Topological Signal Processing and Learning

SPEAKER(S)

ABSTRACT

Paolo Di Lorenzo & Sergio Barbarossa | Sapienza University of Rome, Italy
Elvin Isufi | Delft University of Technology, The Netherlands

Many modern datasets exhibit complex relational structures that cannot be properly described using traditional Euclidean models. In recent years, graph signal processing and geometric deep learning have provided powerful frameworks for processing signals defined on irregular domains such as networks. However, many real-world systems involve higher-order interactions among groups of entities, which cannot be adequately captured by pairwise relations. Examples include multi-agent coordination, social group interactions, biological networks, and infrastructure systems where flows occur along network links. To model these interactions, researchers have recently turned to topological representations such as simplicial complexes and cell complexes, which naturally capture multiway relationships. This tutorial introduces the emerging framework of Topological Signal Processing and Learning, which generalizes graph signal processing and graph neural networks to signals defined over higher-order structures. The tutorial presents the fundamental tools needed to represent, process, and learn from topological signals using concepts from algebraic topology and signal processing. We begin by reviewing relational inductive biases and the foundations of graph signal processing. We then introduce topological representations of data, including simplicial complexes and the associated algebraic operators such as incidence matrices and Hodge Laplacians. Building on these tools, we present signal processing methods for topological signals, including filtering, sampling, representation, and dictionary learning. Finally, we discuss emerging learning architectures based on simplicial convolutions that extend graph neural networks to higher-order domains. The tutorial combines theoretical insights with illustrative applications and recent research results, providing attendees with a clear overview of this rapidly growing research direction.