A5062501594- Advanced Graph Neural Networks
- Neural Networks and Applications
- Topological and Geometric Data Analysis
- Distributed Sensor Networks and Detection Algorithms
- Privacy-Preserving Technologies in Data
- Age of Information Optimization
- Sparse and Compressive Sensing Techniques
- Stochastic Gradient Optimization Techniques
- Image and Signal Denoising Methods
- Advanced Neuroimaging Techniques and Applications
- Energy Harvesting in Wireless Networks
- IoT and Edge/Fog Computing
- Machine Learning and ELM
- Functional Brain Connectivity Studies
- Homotopy and Cohomology in Algebraic Topology
- Wireless Communication Security Techniques
- Microwave Imaging and Scattering Analysis
- Service-Oriented Architecture and Web Services
- Blind Source Separation Techniques
- Advanced Memory and Neural Computing
- Indoor and Outdoor Localization Technologies
- Advanced Numerical Analysis Techniques
- Domain Adaptation and Few-Shot Learning
- Semantic Web and Ontologies
- Advanced Wireless Communication Technologies
Harvard University Press
2025
Sapienza University of Rome
2020-2024
Harvard University
2024
Boston University
2024
University of Pennsylvania
2023
University of Perugia
2019
The aim of this paper is to propose a novel dynamic resource allocation strategy for energy-efficient adaptive federated learning at the wireless network edge, with latency and performance guarantees. We consider set devices collecting local data uploading processed information an edge server, which runs stochastic gradient-based algorithms perform continuous adaptation. Hinging on Lyapunov optimization tools, we dynamically optimize radio parameters (e.g., transmitting devices, transmit...
Since their introduction, graph attention networks achieved outstanding results in representation learning tasks. However, these consider only pairwise relations between features associated to the nodes and then are unable fully exploit higher-order long-range interactions present many real world data-sets. In this paper, we introduce a neural architecture operating on data defined over edges of graph, represented as 1-skeleton regular cell complex, able capture insightful interactions....
Graph Neural Networks (GNNs) excel at learning from graph-structured data but are limited to modeling pairwise interactions, insufficient for capturing higher-order relationships present in many real-world systems. Topological Deep Learning (TDL) has allowed systematic of hierarchical interactions by relying on combinatorial topological spaces such as simplicial complexes. In parallel, Quantum (QNNs) have been introduced leverage quantum mechanics enhanced computational and power. this work,...
This paper introduces topological Slepians, i.e., a novel class of signals defined over spaces (e.g., simplicial complexes) that are maximally concentrated on the domain set nodes, edges, triangles, etc.) and perfectly localized dual frequencies). These obtained as principal eigenvectors matrix built from proper localization operators acting topology frequency domains. Then, we suggest principled procedure to build dictionaries which theoretically provide non-degenerate frames. Finally,...
Weighing the topological domain over which data can be represented and analysed is a key strategy in many signal processing machine learning applications, enabling extraction exploitation of meaningful features their (higher order) relationships. Our goal this paper to present tools for weighted simplicial complexes. Specifically, relying on Hodge Laplacian theory, we propose efficient strategies jointly learn weights complex filters solenoidal, irrotational harmonic components signals...
Learning at the edge is a challenging task from several perspectives, since data must be collected by end devices (e.g. sensors), possibly pre-processed compression), and finally processed remotely to output result of training and/or inference phases. This involves heterogeneous resources, such as radio, computing learning related parameters. In this context, we propose an algorithm that dynamically selects encoding scheme, local uplink radio parameters, remote perform classification with...
Internet of Things (IoT) applications combine sensing, wireless communication, intelligence, and actuation, enabling the interaction among heterogeneous devices that collect process considerable amounts data. However, effectiveness IoT needs to face limitation available resources, including spectrum, energy, computing, learning inference capabilities. This article challenges prevailing approach which prioritizes usage resources in order guarantee perfect recovery, at bit level, data...
This paper describes the 2nd edition of ICML Topological Deep Learning Challenge that was hosted within 2024 ELLIS Workshop on Geometry-grounded Representation and Generative Modeling (GRaM). The challenge focused problem representing data in different discrete topological domains order to bridge gap between (TDL) other types structured datasets (e.g. point clouds, graphs). Specifically, participants were asked design implement liftings, i.e. mappings structures --like hypergraphs, or...
Topological Deep Learning (TDL) has emerged as a paradigm to process and learn from signals defined on higher-order combinatorial topological spaces, such simplicial or cell complexes. Although many complex systems have an asymmetric relational structure, most TDL models forcibly symmetrize these relationships. In this paper, we first introduce novel notion of directionality then design Directed Simplicial Neural Networks (Dir-SNNs) based it. Dir-SNNs are message-passing networks operating...
In this work we introduce a convolution operation over the tangent bundle of Riemannian manifolds exploiting Connection Laplacian operator. We use to define filters and neural networks (TNNs), novel continuous architectures operating on signals, i.e. vector fields manifolds. discretize TNNs both in space time domains, showing that their discrete counterpart is principled variant recently introduced Sheaf Neural Networks. formally prove architecture converges underlying TNN. numerically...
Latent Graph Inference (LGI) relaxed the reliance of Neural Networks (GNNs) on a given graph topology by dynamically learning it. However, most LGI methods assume to have (noisy, incomplete, improvable, ...) input rewire and can solely learn regular topologies. In wake success Topological Deep Learning (TDL), we study Topology (LTI) for higher-order cell complexes (with sparse not topology) describing multi-way interactions between data points. To this aim, introduce Differentiable Cell...
The aim of this paper is to propose a novel dynamic resource allocation strategy for energy-efficient federated learning at the wireless network edge, with latency and performance guarantees. We consider set devices collecting local data uploading processed information an edge server, which runs stochastic gradient descent (SGD) perform distributed adaptation. Hinging on Lyapunov optimization tools, we dynamically optimize radio parameters (i.e., transmitting devices, transmit powers)...
We introduce topox, a Python software suite that provides reliable and user-friendly building blocks for computing machine learning on topological domains extend graphs: hypergraphs, simplicial, cellular, path combinatorial complexes. topox consists of three packages: toponetx facilitates constructing these domains, including working with nodes, edges higher-order cells; topoembedx methods to embed into vector spaces, akin popular graph-based embedding algorithms such as node2vec; topomodelx...
The aim of this work is to propose a novel dynamic resource allocation strategy for adaptive Federated Learning (FL), in the context beyond 5G networks endowed with Reconfigurable Intelligent Surfaces (RISs). Due time-varying wireless channel conditions, communication resources (e.g., set transmitting devices, transmit powers, bits), computation parameters CPU cycles at devices and server) RISs reflectivity must be optimized each round, order strike best trade-off between power, latency,...
The goal of this paper is to introduce pooling strategies for simplicial convolutional neural networks. Inspired by graph methods, we a general formulation layer that performs: i) local aggregation signals; ii) principled selection sampling sets; iii) downsampling and topology adaptation. then customized design four different (i.e., max, top-k, self-attention, separated top-k) grounded in the theory topological signal processing. Also, leverage proposed layers hierarchical architecture...
The aim of this paper is to introduce a novel dictionary learning algorithm for sparse representation signals defined over regular cell complexes. Leveraging tools from Hodge theory, we inject the underlying topology in structure by parametrizing it as concatenation sub-dictionaries that are polynomial Laplacians. problem cast joint optimization topological coefficients and signal representation, which efficiently solved via an iterative alternating algorithm. Numerical results on synthetic...
We study decentralized estimation of time-varying signals at a fusion center, when energy harvesting sensors transmit sampled data over rate-constrained links.We propose dynamic strategies to select radio parameters, sampling set, and harvested each node, with the aim estimating signal while ensuring: i) accuracy recovery procedure, ii) stability batteries around prescribed operating level.The approach is based on stochastic optimization tools, which enable adaptive without need apriori...
We study decentralized estimation of time-varying signals at a fusion center (FC), when energy harvesting sensors transmit sampled data over rate-constrained links. propose dynamic strategy based on stochastic optimization for selecting radio parameters, sampling set, and harvested each node, with the aim estimating signal guaranteed performance while ensuring stability batteries around prescribed operating level. Numerical results validate proposed approach under communication constraints.
This work aims to propose a novel architecture and training strategy for graph convolutional networks (GCN). The proposed architecture, named Autoencoder-Aided GCN (AA-GCN), compresses the features in an information-rich embedding at multiple hidden layers, exploiting presence of autoencoders before point-wise nonlinearities. Then, we end-to-end procedure that learns different representations per layer, jointly with weights auto-encoder parameters. As result, improves computational...
In this work, we study the problem of stability Graph Convolutional Neural Networks (GCNs) under random small perturbations in underlying graph topology, i.e. a limited number insertions or deletions edges. We derive novel bound on expected difference between outputs unperturbed and perturbed GCNs. The proposed explicitly depends magnitude perturbation eigenpairs Laplacian matrix, which edges are inserted deleted. Then, provide quantitative characterization effect perturbing specific...
Topological neural networks (TNNs) are information processing architectures that model representations from data lying over topological spaces (e.g., simplicial or cell complexes) and allow for decentralized implementation through localized communications different neighborhoods. Existing TNN have not yet been considered in realistic communication scenarios, where channel effects typically introduce disturbances such as fading noise. This paper aims to propose a novel design, operating on...