Special sessions

Continual Learning (CL) and Streaming Machine Learning (SML) deal with the adaptation of learning agents in dynamic environments that change over time. However, these two communities have rarely interacted, mostly because they leverage different learning approaches and evaluate performance according to different metrics. While CL mostly focuses on consolidating knowledge over time without forgetting, often irrespective of
real-time requirements, SML focuses on fast and efficient adaptation on a high-frequency stream of data.

This special session aims to connect the CL and SML communities. While we welcome works that push the state of the art in either topic, we strongly encourage novel works at their intersection. For example, applying deep neural networks in SML scenarios could be particularly valuable in contexts where complex patterns or temporal dependencies in the data need to be captured and effectively modeled. Since memory consolidation with such models is heavily studied in CL, many existing approaches may be directly used in SML. Vice versa, the fast adaptation ability of SML models and concept drift detection methodologies can be leveraged to improve the convergence speed and forward transfer of CL. This is particularly interesting for challenging scenarios like Online CL, where the
high-frequency data stream is fully compatible with SML.

Topics include, but are not limited to:

• Continual / Lifelong learning
• Streaming machine learning
• Continual / Streaming Machine Learning benchmarks
• Continual / Streaming Machine Learning libraries
• Applications of Streaming Continual Learning
• Online learning
• Learning in non-stationary environments
• Test-Time adaptation
• Domain adaptation

Graphs are versatile abstractions that model complex systems of inter- acting entities, where interactions imply functional and/or structural depen- dencies among them. Molecular compounds and social networks are two particularly relevant instances of graph-structured data: the former can be viewed as a system of interacting atoms, where chemical bonds are influenced by factors like inter-atomic distances and atomic energies/forces depend on long-range electrostatic interactions. Instead, social networks manifest in terms of user-user and user-content interactions, where content is multimodal and includes pictures, movies, and songs. Other practical examples include the encoding of symmetries and constraints in combinatorial optimization problems, which serve as proxies for apriori knowledge.

The field of deep learning for graph-structured data focuses on adapt- ing deep learning techniques, such as convolutional operators, to the analysis and processing of graphs. Recent advancements in foundation and generative models for non-graph data have significantly impacted businesses, with an unprecedented speed of adoption in production environments. It is therefore natural to ask ourselves whether such advances can be readily transferred to the domain of graphs, and what the best way to do so would be, opening new avenues for research and applications. In this respect, there are high expectations in areas such as drug discovery, where the ability to generate molecular constraint-preserving graphs with desired properties might reduce the humongous amount of money and compute time needed to screen candi- date drugs.

This special session aims to gather valuable contributionsbac and new find- ings in the field of deep learning for graphs. It will also emphasize the integration of foundation and generative models to enhance the creation and manipulation of graph-structured data, unlocking new possibilities in areas such as molecular and material design.

In particular, we look for contributions in the following areas:

• Architectures for foundation models operating on graphs;
• Graph generation (e.g., probabilistic models, variational autoencoders, normalizing flow, diffusion models);
• Graph representation learning;
• Graph structure learning and relational inference;
• Graph coarsening and pooling in graph neural networks;
• Theory of graph neural networks (e.g., expressive power, learnability, negative results);
• Learning on complex graphs (e.g., dynamic graphs, heterogeneous graphs, and spatio-temporal data);
• Anomaly and change detection in graph data;
• Randomized neural networks for graphs (e.g., reservoir computing);
• Recurrent, recursive, and contextual models;
• Scalability, data efficiency, and training techniques of graph neural net- works;
• Tensor methods for structured data;
• Graph datasets and benchmarks.

Graph generation is developing fast as a new research branch for deep learning models, leading to remarkable solutions across various application domains, including Social Network Analysis, Recommendation Systems, Telecommunications and Power Grids, Supply Chain & Logistics, and others.
Graph generators can create graph structures with different characteristics and purposes. This special session is dedicated both to theoretical advancements in algorithms and models that are driving these innovations, as well as practical applications and real-world results, with a particular interest in life sciences.

This special session welcomes contributions that explore various graph generation methodologies in life sciences, such as sequential methods, graph variational autoencoders, and graph diffusion models. A particular emphasis will be devoted to new and promising directions for future exploration in the research community, such as, for example: conditional and hierarchical graph generation techniques, graph diffusion models, very deep networks for graph generation.

Relevant topics include, but are not limited to:

• Drug Discovery & Molecular Design
• Molecular Graph Generation
• Protein Structure Prediction
• Metabolic Networks and Pathways
• Protein-Protein Interactions
• Gene Regulation Networks
• Antibody Generation
• Benchmarks for Graph Generation
• Graph Fusion
• Text to graph diffusion models
• Interpretable, Explainable, Trustworthy graph generation

Longitudinal data arise in a variety of real-life applications where individuals are observed during their lifetime or some given timespan.  Their modelling represents a challenging topic, due to their complexity (temporal, censored, multivariate, possibly irregularly sampled, possibly multimodal information including text, images, …). For example, aging is the human temporal process par excellence, and projects like ARLES aim at exploring Aging Risks and their Long-term impact on the Economy and Society by means of mathematical models.  In this session, we welcome both theoretical and methodological contributions, as well as practical implementations related to or applicable to the analysis of longitudinal data in social and biomedical sciences. Particular interest will be given to proposals focusing on the analysis of stochastic processes for population evolution, heavy tailed distributions, change-point detection, hybridisation with functional data methods, and the use of neural networks for modelling and forecasting these specific data. On the side of applications,  we expect contributions using data from the social or the biomedical fields, like for instance, the design of aging clocks from the statistical analysis of biomarkers. Eventually, through this session we wish to enhance the availability of data and methods in these domains and foster multidisciplinary collaboration among research groups.

On the wake of the finalisation of the European flagship Human Brain Project and similar initiatives around the world, it has become clear that Cognitive Science has become a data-centric endeavour. Cognitive Science is an interdisciplinary research field comprising neuroscience, psychology, linguistics, artificial intelligence, and philosophy. In particular, mental health is an important topic both for Cognitive Science and Healthcare.  Data-based approaches stemming from Artificial Intelligence and particularly from Machine Learning are important tools in the pursuit of personalized medicine goals as means for improvement in diagnosis, prognosis and treatment. In the medical domain in general, both the trustworthiness of models and the explainability of methods are relevant for the adoption of such machine learning-based approaches in clinical diagnosis and treatment. This session will consider developments in the field of applied Artificial Intelligence for Cognitive Science, with a focus on medical applications. Topics of interest are, among others: 

• Neuroscience
• Psychology and Psychiatry
• Neurodegenerative diseases
• Linguistics
• Philosophy
• Artificial Intelligence applications for diagnosis, prognosis and treatment of cognitive impairments and pathologies.
• Patient cohort analysis. 
• Interpretable and explainable artificial intelligence methods in Cognitive Science.  
• Personalized medicine.
• Artificial Intelligence–based medical devices regulation.

The intersection of Network Science and Artificial Intelligence (AI) is a burgeon- ing field that promises to revolutionize our understanding and management of complex systems. This special session, titled ”Network Science Meets AI,” aims to bring together researchers from both disciplines to explore the synergies and innovative applications that arise from their convergence. Network Science has provided a unified framework for representing complex systems through the lens of interactions among elements, often captured by networks that evolve over time and different types of relationships among the same set of elements. This framework has revealed universal properties such as scale-free nature, robust- ness, resilience, and modular structure. It has also deepened our understanding of the dynamics and controllability of complex networks, particularly in areas like contagion processes and epidemic models. On the other hand, AI is transforming various scientific disciplines by introducing new paradigms for scientific discovery. The capabilities of AI to solve highly complex problems, such as protein structure prediction, drug discovery, and content creation through generative AI, are nearing human-level performance.

This special session is aimed at contributions at the intersection between net- work science and AI, where AI offers new tools to tackle problems in network science, or, vice versa, network science supports the design and understanding of AI methods. There are multiple areas where a stronger interaction between network science and AI is promising: For instance, novel paradigms to describe high-order interactions through hypergraphs or simplicial complexes are closely linked to current challenges in the domain of representation learning. This also influences solutions for problems such as label, link, and network property prediction, or for combinatorial optimization problems on networks. Simulating and understanding dynamical processes in complex networks is another key challenge in network science that can benefit from AI, particularly from the capacity to handle non-linear relationships among system elements and non- linearity in the system dynamics. Being able to infer dynamical equations in a data-driven fashion is crucial for the simulation of network dynamics and node/link attribute dynamics along with generating plausible networks related to real-world systems, such as proteins, molecules, transportation systems, or even social networks. Moreover, it is equally important to consider the contribution of network science to advance methods in AI and machine learning. The development of graph neural networks that are based on neural message passing is an important research theme in the deep learning community. Existing insights into the topology of complex networks can greatly assist the design of such deep neural networks for graph-structured data. Moreover, recent works have shown that objective functions for community detection originally devised in the network science community can be repurposed for cluster detection based on deep graph learning. Finally, deep neural networks themselves are becoming instances of complex systems, thus turning into a subject of interest in network science. To this end, network science methods can be used to understand or optimize the structure of deep neural networks. Furthermore, methods for the analysis of dynamics processes can advance the comprehension of the learning processes of the parameters as well as the dynamics of the inference phases. Relevant topics include, but are not limited to:

• Optimizing GNN architectures using Network Science principles
• Novel message passing schemes inspired by complex network dynamics
• Topology-aware neural network design
• Representation learning for hypergraphs and simplicial complexes
• Deep learning models for higher-order interactions
• AI-driven prediction of network evolution
• Deep learning approaches for temporal network embedding
• Neural ODEs for network dynamics
• Explainable graph machine learning algorithms
• Explainability techniques for graph machine learning
• Network Generation and Synthesis
• Topological analysis of deep neural networks
• Transfer of network science objectives to deep learning (e.g. community detection, influence maximization, epidemic modeling)
• Theoretical foundations of graph machine learning methods
• Causal discovery in complex networks
• Topology-aware optimization algorithms

For more details, visit the website: https://sites.google.com/view/esann-netsai/.
 

Quantum physics offers a new paradigm that promises to make certain computations faster and more efficient. The recent progress of quantum computers allows for more com- plex applications which lead to a rising interest in transferring machine learning methods to quantum hardware for practical applications. However, the development of quantum computers is still in its beginnings and currently these approaches requires synergy with classical computers. Moreover, many fundamental questions have not yet been answered. In this session, we would therefore not only look into machine learning on quantum computers themselves, but also into their efficient interplay with classical computers and the open questions that need to be answered on the way to genuine Quantum Machine Learn- ing. These includes topics like classical pre-processing, efficient encoding strategies, and quantum-classical hybrid models, as well as, purely classical machine learning approaches inspired by quantum physics, such as tensor networks.