There will be 3 tutorials at CAMSAP:

1. Training Neural Networks at Any Scale

• 
      Volkan Cevher
      EPFL
      Switzerland
  
  Leena C. Vankadara    
  UCL    
  UK    

Show bios
At the heart of deep learning's transformative impact lies the concept of scale-encompassing both data and computational resources, as well as their interaction with neural network architectures. Scale, however, presents critical challenges, such as increased instability during training and prohibitively expensive model-specific tuning. Given the substantial resources required to train such models, formulating high-confidence scaling hypotheses backed by a rigorous theoretical research has become paramount. The first part of the tutorial will provide an overview of significant advances in the theory of scaling in deep learning, covering its historical foundations, recent breakthroughs, and practical implications for training large-scale models. To bridge theory and practice, the tutorial explores another key mathematical ingredient of scaling: the numerical solution algorithms commonly employed in deep learning, spanning domains from vision to language models. We unify these algorithms under a common master template, making their foundational principles transparent. In doing so, we reveal the interplay between adaptation to smoothness structures via online learning and the exploitation of optimization geometry through non-Euclidean norms. Our exposition moves beyond simply building larger models--it emphasizes strategic scaling, offering insights that promise to advance the field while economizing on resources.
 

2. Tensor Decompositions and Deterministic Blind Signal Separation

• 
      Lieven De Lathauwer
      KU Leuven
      Belgium

   
   

Show bio
Over the years, higher-order tensors have become increasingly important in signal processing, data analysis, and related fields. One of the key reasons for this success is that tensors offer tools for blind signal separation that have no equivalent in matrix-based methods. In particular, the Canonical Polyadic Decomposition (CPD)--the decomposition into rank-1 terms--is unique under relatively mild conditions, enabling the unambiguous recovery of underlying data components. Block Term Decompositions (BTD) build on this success by allowing the unique recovery of components that are not necessarily rank-1. Blind source separation often involves data that is initially represented as a matrix. A first step is therefore to construct a meaningful tensor from the available data. In the context of Independent Component Analysis (ICA), this tensor is typically derived by estimating higher-order statistics or a collection of second-order statistics. The rank-1 structure arises from the assumed statistical independence of the source signals. Statistics and statistical independence are not the only means of deriving a tensor from the given data. Constant modulus is one example of a deterministic property that can be exploited for this purpose, as demonstrated in the Analytical Constant Modulus Algorithm (ACMA). In this tutorial, we will explore several alternative properties that can be leveraged for deterministic blind source separation. One such property is exponential polynomial structure: if signals can be approximately modeled as sums and/or products of exponentials, sinusoids, and/or polynomials, this structure can serve as a basis for blind separation. In this case, the tensor is (implicitly) derived from a Hankel representation. Rational function-type structure can also enable blind separation, with the corresponding tensor (implicitly) derived from a Loewner representation. In large-scale settings, segmentation (or quantization) may be employed. We will highlight several (often hidden) properties of time series that can be translated into low-rank structure. Although the exposition is grounded in solid mathematical principles, the emphasis will be on clarifying the reasoning behind the techniques and building intuition. The discussion of foundational topics such as CPD will include recent developments, such as uniqueness results in noisy settings. Concepts will be illustrated through use cases and application examples.
 

3. Graph-based Machine Learning for Wireless Communications

• 
      Santiago Segarra
      Rice University
      USA

   
   

Show bio
As communication networks continue to grow in scale and complexity, traditional approaches to network design and operation are becoming inadequate. Machine learning (ML) has garnered significant attention for its potential to complement conventional mathematical models in the capabilities of describing complex wireless systems and deriving computationally efficient solutions. However, standard ML methods, such as multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs), struggle to effectively leverage the underlying topology of communication networks, causing significant performance degradation as network size increases. Graph neural networks (GNNs) emerge as a promising ML approach that has gained surging interest within the ML community. GNNs excel when dealing with large network scales and dynamic topologies, outperforming MLPs and CNNs in such scenarios. This timely tutorial provides a gentle introduction to GNNs and explores recent approaches for applying them to solve classical problems in wireless communications and networking, including power allocation, beamforming, link scheduling, and routing. The emphasis will be on how GNNs can augment, rather than replace, existing solutions to these problems. The goal of the tutorial is to foster further research and exchange between the communications, ML, and signal processing communities and to inspire applications of ML in fields that will further advance wireless communications.