Energy efficiency in the Information and Communication Technologies (ICT) ecosystem has always been a pillar for reducing the energy used and enabling higher integration. However, despite the continuous increment in efficiency, the energy consumption of ICT has kept increasing during the last decades and it now starts to become a pressing problem that we urgently need to solve. To keep the temperature increment below 2ºC we need to cut the CO2 emissions by half by 2030, which means a decrease of 7% per year on a world scale basis. ICT is part of the problem: recent estimations fix its contribution to global emissions high as 2.1%–3.9% [1]. Moreover, ICT is one of the sectors that is experiencing the highest growth in terms of CO2 emissions, reaching 6% nowadays, and, with the current trend, it will be eating up 15% by 2030. In addition, new services based on AI are becoming more pervasive each day, with models that are exponentially increasing in their complexity. A very recent example from standard Centralized AI (CAI) can be found in OpenAI´s ChatGPT service, whose monthly electricity consumption is estimated to be millions of kWh: “ChatGPT may have consumed as much electricity as 175,000 people in January 2023" [2]. Similarly, Meta used 2.6 million KWh hours of electricity and emitted 1,000 tons of CO2 for the development of their new Large Language Model Meta AI (LLaMA) models [3]. This energy is estimated to have led to the emission of 1,015 tCO2e, roughly the annual carbon footprint of 150 EU citizens [4]. Moreover, CAI suffers from managing data coming from distributed sources in a centralized fashion. In fact, data must be transmitted to the cloud to be processed, with the consequent issues of high network congestion, waste of energy, and high latency. Furthermore, cloud data centers are normally managed by a third party, which may represent a breach in the privacy preservation of sensitive content. The Edge Intelligence (EI) paradigm has emerged as a mitigation to the challenges described above. EI involves performing AI computations directly at the network edge, where most of the data is generated. This approach offers several advantages with respect to CAI, e.g. reduced communication overhead, faster response time, and increased privacy preservation. Current research in EI focuses on distributing learning processes to achieve accuracy levels comparable to CAI [5], regardless of the energy demands of computation and communication. However, considering the exponential upsurge of data generated [6], the continued expansion of connected devices, and the projected AI annual growth rate of 37.3% from 2023 to 2030 [7], incorporating sustainable design principles into EI platforms is essential to prevent environmental degradation and avoid the rebound effect problem [1]. This shift in AI design and application is the central objective of the PSI project. 

PSI proposes a novel energy-aware integrated computing and communication platform that will radically transform the current energy-intensive AI paradigm by centralized CAI. In this framework, traditional data centers will be replaced by far-edge devices, such as cyber-physical systems (CPSs), becoming key enablers of pervasive AI services, also called Artificial Intelligence of Things (AIoT). These devices, often limited in connectivity, computing capabilities, memory, and reliant on batteries, will require the adoption of highly efficient energy-saving, communication, and computational strategies. Moreover, data distribution across CPSs may exhibit varying statistical characteristics, even when sourced from the same physical system [8], leading to non-identical and independent distributions (non-iid), which can negatively impact the performance of traditional ML models. The PSI proposed solution empowers CPSs to compute accurate ML models collaboratively on-device, reducing or eliminating the reliance on cloud data centers. To overcome the limitations of standard CPSs (e.g., sensor nodes), and their constrained computational and energy resources, highly efficient learning architectures are essential. Existing CAI-based learning solutions are excessively complex for these devices, necessitating the development of novel, streamlined algorithms.  

By implementing decentralized techniques, PSI aims to enable direct communication among devices for both training and inference, removing the need for centralized entities and reducing long-range data transmissions. This approach will mitigate potential bottlenecks in local edge nodes and aggregators, enhancing both system scalability and energy efficiency. The same decentralized and energy-efficient AI architecture will be applicable to a wide range of AIoT contexts, including cognitive buildings, where resource-constrained sensors can manage real-time data for energy-efficient building operations. Moreover, continual learning approaches will ensure that models adapt to non-stationary environments, a typical characteristic in real-world conditions. 

PSI proposes two primary technological enablers to achieve this vision: 

Brain-inspired computing: PSI will consider both reservoir and neuromorphic computing to enable lightweight ML models. Reservoir computing is especially well-suited for learning dynamical systems from time-series data, requiring only small training datasets, using linear optimization, and demanding minimal computational resources [9]. This makes it a natural fit for energy-limited CPSs. In neuromorphic computing circuits are made of physical neurons interconnected by physical synapses with in-situ, non-volatile memory, drastically reducing the need for data transfer within the circuit. Thus, it offers substantial improvements in speed and energy efficiency [10], making it a key driver for edge intelligence. The interplay between the reservoir and neuromorphic computing will also be considered to further improve energy savings. 

Distributed and collaborative learning: Techniques such as Federated Learning (FL) and Continual Learning (CL) [11][12][13] enable efficient model sharing across data sources, fostering knowledge distillation and enhancing the system´s generalization capabilities, especially in the presence of non-iid data. These approaches are essential for achieving high-performance ML in distributed environments with limited computational and energy resources. In particular, FL aims to coordinate the nodes to learn from their local data simultaneously and merge the shared knowledge in a parallel fashion. Whereas CL can be used to learn from the distributed source of knowledge sequentially, hybrid solutions that combine the strengths of both FL and CL must be explored to create a truly sustainable AI paradigm tailored to specific applications.  

It is important to note that, to date, these areas have not yet been thoroughly studied from an energy-aware perspective, nor have lightweight ML models (such as those based on reservoir or neuromorphic computing) been considered in distributed learning systems. This represents a significant gap in the state-of-the-art, one that this project seeks to fill. 

In summary, PSI advocates for a pervasive and liquid AI system, where edge devices can actively contribute to AI service provision, the AIoT [14]. Through this approach, CPSs will collaboratively process data from the underlying physical systems using lightweight, brain-inspired ML models, regardless of their resource capabilities. This paradigm will position users and devices as active participants in the AI process, not merely passive data generators. Moreover, it has the potential to democratize AI, shifting control away from large tech companies and enabling more transparent, user-centered AI systems. Finally, this approach also enhances user privacy by keeping personal data at the edge, avoiding the need to transfer it to third-party data centers with limited transparency. 

PSI will apply novel research methodologies to the sustainable design of AIoT devices, platforms, and services, through the interdisciplinary understanding of this timely technology area of key importance for society. In particular, the aim is to persuade that the current narrative that “the larger the AI system the more valuable, powerful, and interesting” can be changed by working with small and collaborative ML solutions. In the last decade, the research community embraced the bigger-is-the-better AI paradigm with the consequence of an explosion in investment in large-scale AI models [15]. From a research perspective, reviewers ask for experiments at a large scale, both in the context of new models and in the context of measuring performance against existing models, which implies running experiments many times to search for the optimal hyperparameter. This pushes the need for computing capabilities beyond the common budgets of most university labs and, in turn, many labs have to be increasingly dependent on industry to secure such access. The final consequence is that investigating AI can be increasingly difficult for anyone outside of large industrial labs. The expense of the computing elements needed to build and operate large AI models is benefiting the actors in AI that have already access to it. This phenomenon concentrates power and influence over AI, perpetuating the bigger-is-better AI paradigm in service of maintaining their market advantage. An example can be found in the recent shortages in chips when the gap between the “GPU rich” and “GPU poor” has increased due to the lack of computing resources. However, choosing the right model architecture for the data at hand is crucial, and Transformer-based models, widely perceived to be SOTA on most or even all ML benchmarks, are not always the most fitting solution. For this, it is useful to consider the trade-off between task performance (e.g. performance on an ML benchmark) and the computing resources used, the so-called Green AI paradigm. PSI wants to persuade that task performance can still play an important role and compete with bigger-is-better AI, especially when considering intrinsic distributed problems, as for AIoT. To do so, PSI wants to show the limitations and expectations of brain-inspired ML solutions when working in a collaborative environment. This would open a more democratic paradigm for doing research on AI, enabling the use of small-scale and less expensive hardware. 

The remarkable success of attention-based models in real-world applications has sparked a crucial question for Reservoir Computing (RC): Can its inherent computational efficiency compete with the high-performance, yet energy-intensive, novel deep learning architectures? Can Deep and modular RC neural networks address state-of-the-art challenges in Computer Vision and Natural Language Processing? In the attempt to consolidate RC capabilities towards more complex tasks, this paper delves into the exploration of a comprehensive RC’s offline-
online cycle cost analysis. Our investigation highlights hyper-parameters (HPs) optimization as a major bottleneck in RC deployment, particularly for those exploring RC capabilities and those who want to maintain user-level knowledge of the solution. To address this, we introduce an adaptive ϵ-Greedy based search exploration mechanism, significantly streamlining the off-line optimization process while maintaining high accuracy. Furthermore, we enhance existing RC frameworks to support online transfer learning and inference, enabling seamless, fast, and energy-efficient adaptation to real-world environments. By analyzing the impact of optimized HPs on performance, we aim to demonstrate the viability of RC as a powerful and efficient alternative for many practical applications, including those on devices with limited resources. Experimental results proved that our solution is able to reduce the time required for offline HPs optimization by 70%, enabling energy savings of up to
88%. Moreover, in the online scenario, it guarantees similar performance in terms of accuracy while reducing memory usage by 66%.

You may find the full paper the following link.

The rise of Generative AI has renewed interest in Deep Learning across academia and industry, attracting smaller research groups and non-IT companies eager to leverage Machine Learning (ML). However, high infrastructure costs often make AI adoption impractical. Democratizing access to High-Performance Computing (HPC) is key to overcoming this barrier, enabling broader ML adoption while reducing e-waste by integrating existing resources. DARE-ML (Democratized Accessible Resource Environment for Machine Learning) offers a framework to optimize resource allocation, lower energy use, and improve ML accessibility. By profiling models interactively in a heterogeneous, limited-GPU environment, DARE-ML collects key metadata—like training time and memory needs—before scheduling jobs. At its core, DARE-ML incorporates an efficient interactive profiling mechanism powered by ESN (Echo State Networks), enabling streamlined and resource-aware execution of deep learning tasks. Experiments in real and simulated settings show DARE-ML improves ML job scheduling, reducing energy use up to 80-fold and cutting both average job completion time and waiting time per user by 15%.
You may find the full paper at the following link.

The growing complexity of Cyber-Physical Systems (CPS) in industrial and manufacturing environments calls for more sophisticated methods to represent heterogeneous assets and processes. In response, hierarchical Digital Twins (DTs)—virtual representations of physical, taxonomy-based processes—offer transparent, layered modeling of diverse data sources. This layered structure fuels renewed interest in intelligent engines capable of extracting meaningful insights and mapping them within the stratified DT ecosystem. While current Intelligent Digital Twin (I-DT) engines based on Deep Learning are computationally demanding, lightweight alternatives like Reservoir Computing (RC) offer efficient solutions with low training costs and fast inference for modeling causal dynamics. This inherent trade-off between performance and practicality underscores the limitations of evaluating I-DTs on accuracy alone. To address this gap, this work introduces a novel metric, Fidelity, designed to provide a comprehensive evaluation. Unlike traditional approaches, Fidelity also accounts for maintainability and deployability, especially in contexts involving time-varying and hierarchical data dynamics. Extensive experiments on two multimodal datasets demonstrate the competitiveness of our RC-based engine and highlight the value of introducing Fidelity for effectively profiling I-DTs. Specifically, our RC-based engine, identified as optimal through a higher Fidelity score, consumes an order of magnitude less energy and achieves up to 39% higher accuracy (about 10% increase on average) compared to both canonical and other RC-based alternatives.

You may find the full paper at the following link.

[1] C. Freitag, M. Berners-Lee, K. Widdicks, B. Knowles, G. S. Blair, A. Friday, “The real climate and transformative impact of ICT: A critique of estimates, trends, and regulations”, Patterns, Volume 2, Issue 9, 2021, 100340, ISSN 2666-3899, https://doi.org/10.1016/j.patter.2021.100340. 

[2] ChatGPT’s Electricity Consumption, Kasper Groes Albin Ludvigsen, Toward Data Science,  https://towardsdatascience.com/chatgpts-electricity-consumption-7873483feac4 

[3] H. Touvron et al  “LLaMA: Open and Efficient Foundation Language Models”, 2023, ArXiv, abs/2302.13971. 

[4] Facebook disclose the carbon footprint of their new LLaMA models, Kasper Groes Albin Ludvigsen, Medium, https://kaspergroesludvigsen.medium.com/facebook-disclose-the-carbon-footprint-of-their-new-llama-models-9629a3c5c28b 

[5] Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang, ‘‘Edge intelligence: Paving the last mile of artificial intelligence with edge computing,’’ Proc. IEEE, vol. 107, no. 8, pp. 1738–1762, Aug. 2019 

[6] ACEEE. 2017. Intelligent Efficiency Technology and Market Assessment. Report IE1701. Prepared by E.A. Rogers and E. Junga of the American Council for an Energy Efficient Economy. 

[7] Forbes, “24 Top AI Statistics And Trends In 2024”, https://www.forbes.com/advisor/business/ai-statistics/#sources_section 

[8] Zhu, J. Xu, S. Liu, Y. Jin, “Federated learning on non-IID data: A survey”, Neurocomputing, Volume 465, 2021, Pages 371-390,  ISSN 0925-2312,  https://doi.org/10.1016/j.neucom.2021.07.098. 

[9] D.J. Gauthier, E. Bollt, A. Griffith, et al. Next generation reservoir computing. Nat Commun12, 5564 (2021). https://doi.org/10.1038/s41467-021-25801-2 

[10] Big data needs a hardware revolution. Nature http://www.nature.com/articles/d41586-018-01683-1 (2018) doi:10.1038/d41586-018-01683-1. 

[11] M. Miozzo, Z. Ali, L. Giupponi and P. Dini, "Distributed and Multi-Task Learning at the Edge for Energy Efficient Radio Access Networks," in IEEE Access, vol. 9, pp. 12491-12505, 2021, doi: 10.1109/ACCESS.2021.3050841. 

[12] C. Lanza, E. Angelats, M. Miozzo and P. Dini, "Urban Traffic Forecasting using Federated and Continual Learning," 2023 6th Conference on Cloud and Internet of Things (CIoT), Lisbon, Portugal, 2023, pp. 1-8, doi: 10.1109/CIoT57267.2023.10084875. 

[13] E. Guerra, F. Wilhelmi, M. Miozzo and P. Dini, "The Cost of Training Machine Learning Models Over Distributed Data Sources," in IEEE Open Journal of the Communications Society, vol. 4, pp. 1111-1126, 2023, doi: 10.1109/OJCOMS.2023.3274394. 

[14] Zhang and D. Tao, "Empowering Things With Intelligence: A Survey of the Progress, Challenges, and Opportunities in Artificial Intelligence of Things," in IEEE Internet of Things Journal, vol. 8, no. 10, pp. 7789-7817, 15 May 15, 2021. 

[15] Varoquaux, G., Luccioni, A. S., & Whittaker, M. (2024). Hype, Sustainability, and the Price of the Bigger-is-Better Paradigm in AI. arXiv preprint arXiv:2409.14160. 

[16] M. Mendula, C. Leonelli, M. Miozzo, P. Dini “DARE-ML: Democratized Accessible Resource Environment for Machine Learning in the SUPERCOM platform”, accepted for publication at EESP Workshop, held in conjunction with ISC High Performance 2025, Hamburg, June 2025, [pdf].

[17] M. Mendula, M. Miozzo and P. Dini, "Reservoir Computing in Real-World Environments: Optimizing the Cost of Offline and Online Training", accepted at Reservoir Computing in the Deep Learning era:  theory, models, applications, and hardware implementations - Special Session of the International Joint Conference on Neural Networks (IJCNN), Rome July, 2025. [pdf]

[18] M. Mendula, M. Miozzo, P. Bellavista, P. Dini, Reservoir Computing for Enhanced Fidelity in Hierarchical Digital Twin Ecosystems, Future Generation Computer Systems, 2025, 108146, ISSN 0167-739X, https://doi.org/10.1016/j.future.2025.108146 [pdf]