A5011604061- Neural Networks and Reservoir Computing
- Advanced Memory and Neural Computing
- Neural Networks and Applications
- Neural dynamics and brain function
- Machine Learning and ELM
- Model Reduction and Neural Networks
- Advanced Graph Neural Networks
- Optical Network Technologies
- Distributed Control Multi-Agent Systems
- IoT and Edge/Fog Computing
- Context-Aware Activity Recognition Systems
- Data Stream Mining Techniques
- Nonlinear Dynamics and Pattern Formation
- Robotics and Automated Systems
- Anomaly Detection Techniques and Applications
- Modular Robots and Swarm Intelligence
- Ferroelectric and Negative Capacitance Devices
- Domain Adaptation and Few-Shot Learning
- Complex Network Analysis Techniques
- Advanced Adaptive Filtering Techniques
- Autonomous Vehicle Technology and Safety
- Human-Automation Interaction and Safety
- Machine Learning in Healthcare
- Adversarial Robustness in Machine Learning
- Human Mobility and Location-Based Analysis
University of Pisa
2016-2025
Abstract Recent years have witnessed a surge of interest in learning representations graph-structured data, with applications from social networks to drug discovery. However, graph neural networks, the machine models for handling face significant challenges when running on conventional digital hardware, including slowdown Moore’s law due transistor scaling limits and von Neumann bottleneck incurred by physically separated memory processing units, as well high training cost. Here we present...
In this paper we introduce the Graph Echo State Network (GraphESN) model, a generalization of (ESN) approach to graph domains. GraphESNs allow for an efficient Recursive Neural Networks (RecNNs) modeling extended deal with cyclic/acyclic, directed/undirected, labeled graphs. The recurrent reservoir network computes fixed contractive encoding function over graphs and is left untrained after initialization, while feed-forward readout implements adaptive linear output function. Contractivity...
Activity recognition plays a key role in providing activity assistance and care for users smart homes. In this work, we present an system that classifies the near real-time set of common daily activities exploiting both data sampled by sensors embedded smartpho ne carried out user reciprocal Received Signal Strength (RSS) values coming from worn wireless sensor devices deployed environment. order to achieve effective responsive classification, decision tree based on multisensor data-stream...
We address the efficiency issue for construction of a deep graph neural network (GNN). The approach exploits idea representing each input as fixed point dynamical system (implemented through recurrent network), and leverages architectural organization units. Efficiency is gained by many aspects, including use small very sparse networks, where weights units are left untrained under stability condition introduced in this work. This can be viewed way to study intrinsic power architecture GNN,...
In this work, we introduce a novel computational framework inspired by the physics of memristive devices and systems, which embed into context Recurrent Neural Networks (RNNs) for time-series processing. Our proposed memristive-friendly neural network architecture leverages both principles Reservoir Computing (RC) fully trainable RNNs, providing versatile platform sequence learning. We provide mathematical analysis stability resulting dynamics, identifying role crucial RC-based architectural...
The study of deep recurrent neural networks (RNNs) and, in particular, Reservoir Computing (RC) is gaining an increasing research attention the community. recently introduced Deep Echo State Network (DeepESN) model opened way to extremely efficient approach for designing temporal data. At same time, DeepESNs allowed shed light on intrinsic properties state dynamics developed by hierarchical compositions layers, i.e. bias depth RNNs architectural design. In this paper, we summarize...
This work proposes a first study, through empirical assessment, of deep recursive Neural Network (RecNN) architecture for tree structured data exploiting the efficient design Echo State (ESN) framework. Three benchmark tasks trees allow us to assess potentiality novel Deep Tree ESN (DeepTESN) model with respect shallow counterpart (Tree ESN) and literature results (including hidden Markov models kernel based approaches) in different conditions according both efficiency predictive performance.
Echo state networks (ESNs) are time series processing models working under the echo property (ESP) principle. The ESP is a notion of stability that imposes an asymptotic fading memory input. On other hand, resulting inherent architectural bias ESNs may lead to excessive loss information, which in turn harms performance certain tasks with long short-term requirements. To bring together and ability retain as much possible, this article, we introduce new ESN architecture called Edge Stability...
Graph State Space Models (SSMs) have recently been introduced to enhance Neural Networks (GNNs) in modeling long-range interactions. Despite their success, existing methods either compromise on permutation equivariance or limit focus pairwise interactions rather than sequences. Building the connection between Autoregressive Moving Average (ARMA) and SSM, this paper, we introduce GRAMA, a Adaptive method based learnable framework that addresses these limitations. By transforming from static...
Graph Neural Networks (GNNs) are models that leverage the graph structure to transmit information between nodes, typically through message-passing operation. While widely successful, this approach is well known suffer from over-smoothing and over-squashing phenomena, which result in representational collapse as number of layers increases insensitivity contained at distant poorly connected respectively. In paper, we present a unified view these problems lens vanishing gradients, using ideas...
In this paper, we introduce a novel architecture for conditionally activated neural networks combining hierarchical construction of multiple Mixture Experts (MoEs) layers with sampling mechanism that progressively converges to an optimized configuration expert activation. This methodology enables the dynamic unfolding network's architecture, facilitating efficient path-specific training. Experimental results demonstrate approach achieves competitive accuracy compared conventional baselines...
Arctic sea ice, the vast body of frozen water near North Pole, has been in steady decline since satellite observations began. While state-of-the-art models attempt to project future scenarios, they often show significant discrepancies, even though ice system is generally considered linearly with rising temperatures. Machine learning models, although may lack ability fully explain underlying physical processes, offer a complementary approach. By training these on existing data, we can...
A common problem in Message-Passing Neural Networks is oversquashing -- the limited ability to facilitate effective information flow between distant nodes. Oversquashing attributed exponential decay transmission as node distances increase. This paper introduces a novel perspective address oversquashing, leveraging dynamical systems properties of global and local non-dissipativity, that enable maintenance constant rate. We present SWAN, uniquely parameterized GNN model with antisymmetry both...