-
Joint Diffusion Processes as an Inductive Bias in Sheaf Neural Networks
Authors:
Ferran Hernandez Caralt,
Guillermo Bernárdez Gil,
Iulia Duta,
Pietro Liò,
Eduard Alarcón Cot
Abstract:
Sheaf Neural Networks (SNNs) naturally extend Graph Neural Networks (GNNs) by endowing a cellular sheaf over the graph, equipping nodes and edges with vector spaces and defining linear mappings between them. While the attached geometric structure has proven to be useful in analyzing heterophily and oversmoothing, so far the methods by which the sheaf is computed do not always guarantee a good perf…
▽ More
Sheaf Neural Networks (SNNs) naturally extend Graph Neural Networks (GNNs) by endowing a cellular sheaf over the graph, equipping nodes and edges with vector spaces and defining linear mappings between them. While the attached geometric structure has proven to be useful in analyzing heterophily and oversmoothing, so far the methods by which the sheaf is computed do not always guarantee a good performance in such settings. In this work, drawing inspiration from opinion dynamics concepts, we propose two novel sheaf learning approaches that (i) provide a more intuitive understanding of the involved structure maps, (ii) introduce a useful inductive bias for heterophily and oversmoothing, and (iii) infer the sheaf in a way that does not scale with the number of features, thus using fewer learnable parameters than existing methods. In our evaluation, we show the limitations of the real-world benchmarks used so far on SNNs, and design a new synthetic task -- leveraging the symmetries of n-dimensional ellipsoids -- that enables us to better assess the strengths and weaknesses of sheaf-based models. Our extensive experimentation on these novel datasets reveals valuable insights into the scenarios and contexts where SNNs in general -- and our proposed approaches in particular -- can be beneficial.
△ Less
Submitted 30 July, 2024;
originally announced July 2024.
-
Wearable uBrain: Fabric Based-Spiking Neural Network
Authors:
Frances Cleary,
Witawas Srisa-an,
Beatriz Gil,
Jaideep Kesavan,
Tobias Engel,
David C. Henshall,
Sasitharan Balasubramaniam
Abstract:
On garment intelligence influenced by artificial neural networks and neuromorphic computing is emerging as a research direction in the e-textile sector. In particular, bio inspired Spiking Neural Networks mimicking the workings of the brain show promise in recent ICT research applications. Taking such technological advancements and new research directions driving forward the next generation of e-t…
▽ More
On garment intelligence influenced by artificial neural networks and neuromorphic computing is emerging as a research direction in the e-textile sector. In particular, bio inspired Spiking Neural Networks mimicking the workings of the brain show promise in recent ICT research applications. Taking such technological advancements and new research directions driving forward the next generation of e-textiles and smart materials, we present a wearable micro Brain capable of event driven artificial spiking neural network computation in a fabric based environment. We demonstrate a wearable Brain SNN prototype with multi-layer computation, enabling scalability and flexibility in terms of modifications for hidden layers to be augmented to the network. The wearable micro Brain provides a low size, weight and power artificial on-garment intelligent wearable solution with embedded functionality enabling offline adaptive learning through the provision of interchangeable resistor synaptic weightings. The prototype has been evaluated for fault tolerance, where we have determine the robustness of the circuit when certain parts are damaged. Validations were also conducted for movements to determine if the circuit can still perform accurate computation.
△ Less
Submitted 25 February, 2022;
originally announced February 2022.
-
Body Sensor Network: A Self-Adaptive System Exemplar in the Healthcare Domain
Authors:
Eric Bernd Gil,
Ricardo Caldas,
Arthur Rodrigues,
Gabriel Levi Gomes da Silva,
Genaína Nunes Rodrigues,
Patrizio Pelliccione
Abstract:
Recent worldwide events shed light on the need of human-centered systems engineering in the healthcare domain. These systems must be prepared to evolve quickly but safely, according to unpredicted environments and ever-changing pathogens that spread ruthlessly. Such scenarios suffocate hospitals' infrastructure and disable healthcare systems that are not prepared to deal with unpredicted environme…
▽ More
Recent worldwide events shed light on the need of human-centered systems engineering in the healthcare domain. These systems must be prepared to evolve quickly but safely, according to unpredicted environments and ever-changing pathogens that spread ruthlessly. Such scenarios suffocate hospitals' infrastructure and disable healthcare systems that are not prepared to deal with unpredicted environments without costly re-engineering. In the face of these challenges, we offer the SA-BSN -- Self-Adaptive Body Sensor Network -- prototype to explore the rather dynamic patient's health status monitoring. The exemplar is focused on self-adaptation and comes with scenarios that hinder an interplay between system reliability and battery consumption that is available after each execution. Also, we provide: (i) a noise injection mechanism, (ii) file-based patient profiles' configuration, (iii) six healthcare sensor simulations, and (iv) an extensible/reusable controller implementation for self-adaptation. The artifact is implemented in ROS (Robot Operating System), which embraces principles such as ease of use and relies on an active open source community support.
△ Less
Submitted 27 March, 2021;
originally announced March 2021.
-
A Hybrid Approach Combining Control Theory and AI for Engineering Self-Adaptive Systems
Authors:
Ricardo Diniz Caldas,
Arthur Rodrigues,
Eric Bernd Gil,
Genaína Nunes Rodrigues,
Thomas Vogel,
Patrizio Pelliccione
Abstract:
Control theoretical techniques have been successfully adopted as methods for self-adaptive systems design to provide formal guarantees about the effectiveness and robustness of adaptation mechanisms. However, the computational effort to obtain guarantees poses severe constraints when it comes to dynamic adaptation. In order to solve these limitations, in this paper, we propose a hybrid approach co…
▽ More
Control theoretical techniques have been successfully adopted as methods for self-adaptive systems design to provide formal guarantees about the effectiveness and robustness of adaptation mechanisms. However, the computational effort to obtain guarantees poses severe constraints when it comes to dynamic adaptation. In order to solve these limitations, in this paper, we propose a hybrid approach combining software engineering, control theory, and AI to design for software self-adaptation. Our solution proposes a hierarchical and dynamic system manager with performance tuning. Due to the gap between high-level requirements specification and the internal knob behavior of the managed system, a hierarchically composed components architecture seek the separation of concerns towards a dynamic solution. Therefore, a two-layered adaptive manager was designed to satisfy the software requirements with parameters optimization through regression analysis and evolutionary meta-heuristic. The optimization relies on the collection and processing of performance, effectiveness, and robustness metrics w.r.t control theoretical metrics at the offline and online stages. We evaluate our work with a prototype of the Body Sensor Network (BSN) in the healthcare domain, which is largely used as a demonstrator by the community. The BSN was implemented under the Robot Operating System (ROS) architecture, and concerns about the system dependability are taken as adaptation goals. Our results reinforce the necessity of performing well on such a safety-critical domain and contribute with substantial evidence on how hybrid approaches that combine control and AI-based techniques for engineering self-adaptive systems can provide effective adaptation.
△ Less
Submitted 24 April, 2020;
originally announced April 2020.
