Home
Soluzioni
ProgettiPubblicazioniCorsiChi SiamoContatti

Pubblicazioni

scientifiche

  • Categorie

Two-stage dynamic management in energy communities using a decision system based on elastic net regularization
A two-stage forecasting-optimization system for efficient management of energy communities.
A. Rosato, M. Panella, A. Andreotti, Osama A. Mohammed, R. Araneo,
Leggi il paper

Two-stage dynamic management in energy communities using a decision system based on elastic net regularization

A. Rosato, M. Panella, A. Andreotti, Osama A. Mohammed, R. Araneo,
The future of distributed energy: how AI is revolutionizing energy communities
Energy communities are undergoing a groundbreaking transformation. With the increasing adoption of renewable energy sources and energy storage systems, integrating advanced technologies is essential to address challenges such as managing production variability and balancing consumption. Here, artificial intelligence proves indispensable, offering innovative solutions to enhance efficiency and sustainability.
Predicting the future: LSTM networks to anticipate energy demand and supply
Recurrent neural networks, such as LSTMs, enable precise forecasting of renewable energy production and load consumption. This approach anticipates seasonal and daily variations, improving the operational management of distributed resources. A practical example: in an energy community in Southern Italy, which includes 11 loads and 3 distributed generators such as wind turbines and photovoltaic systems, consumption and production profiles were predicted with an average error margin of less than 10%. Simulations included complex scenarios, such as variations in photovoltaic generation during winter, demonstrating that even under adverse weather conditions, the model quickly adapts while maintaining reliable forecasts.
Elastic Net: dynamic optimization to balance consumption and generation
The integration of the Elastic Net methodology in energy management has introduced unprecedented efficiency levels. Using regularization parameters that balance accuracy and efficiency, the system optimizes the operation of storage systems, generators, and loads. In tests conducted in June and December, cumulative imbalances were reduced from 11.5 MWh to 5.5 MWh and from 12.1 MWh to 10.3 MWh, respectively. This result was achieved through careful load management without compromising user comfort. Additionally, batteries were strategically deployed to minimize sudden variations in the grid. The optimization system limits excessive battery use, reducing the risk of premature wear and ensuring optimal operational lifespan.
Real-world applications: success stories in energy communities
The results are tangible: in tests conducted on real data, the system significantly reduced grid imbalances by using storage systems strategically and minimizing reliance on unsustainable energy sources. Moreover, thanks to short-term forecasting, resources such as wind turbines, photovoltaic systems, and electric vehicle charging stations were effectively integrated. This allowed users to actively participate in energy management through economic incentives and reduced operational costs.
A model for sustainable and shared energy
This technology not only enhances the resilience of local grids but also serves as a replicable paradigm in both urban and rural contexts. The combination of AI, Elastic Net, and LSTM networks offers a scalable model for the future of smart grids, where sustainability and innovation go hand in hand.
A Fast Deep Learning Technique for Wi-Fi-Based Human Activity Recognition
A fast AI-based approach for Wi-Fi-based human activity recognition achieves real-time, non-invasive monitoring.
F. Succetti, A. Rosato, F. Di Luzio, A. Ceschini, M. Panella
Leggi il paper

A Fast Deep Learning Technique for Wi-Fi-Based Human Activity Recognition

F. Succetti, A. Rosato, F. Di Luzio, A. Ceschini, M. Panella
The Role of Wi-Fi in Human Activity Recognition
Human Activity Recognition (HAR) is one of the most interesting and promising challenges in the field of artificial intelligence. The applications are numerous, ranging from healthcare to security, with the goal of monitoring and understanding human behavior in real time. Traditionally, this type of analysis has relied on wearable sensors or external devices, but with the pervasive adoption of Wi-Fi networks, a new perspective has emerged. By using Channel State Information (CSI) data from Wi-Fi devices, it is possible to accurately recognize human activities in a non-invasive manner, leveraging Deep Learning techniques for time-series analysis.
AI Technologies for Activity Recognition with CSI
CSI, which contains information about the attenuation and phase shifts of electromagnetic waves during Wi-Fi transmission, has proven to be a much more accurate alternative to other methods such as Received Signal Strength (RSS). Unlike RSS, CSI offers a much more detailed view of movements, able to capture very subtle variations caused by small shifts or changes in human behavior. With advanced deep learning techniques like Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), this data can be analyzed in real-time to recognize a wide range of activities, from walking, sitting, standing, to fall detection.
An Innovative Approach: Deep Neural Networks for Fast Recognition
The introduction of fast Deep Learning models has radically changed how CSI data are analyzed. A practical example of this innovation is the use of a 1D-CNN, which allows for efficient extraction of relevant features, maintaining high accuracy while being fast. This architecture avoids the long training times required by recurrent networks like LSTMs, without sacrificing accuracy in activity classification. In scenarios where fast detection is crucial, this approach marks a significant step forward for real-time applications such as elderly monitoring or smart homes.
Real-World Applications and the Advantages of Wi-Fi-Based Systems
The adoption of Wi-Fi technology for human activity recognition offers numerous tangible benefits. In health monitoring, for example, CSI analysis can be used to detect critical activities such as falls, providing timely assistance without the need for wearable devices. This approach addresses privacy concerns, as it does not require physical contact with the subject being monitored, unlike traditional wearable sensors. Additionally, the system can be easily integrated into existing Wi-Fi infrastructures, reducing costs compared to more invasive solutions like radar or infrared sensors.
A Barrier-Free Future: Non-Invasive Monitoring with AI
Wi-Fi and AI-based activity recognition not only enhances the effectiveness of monitoring solutions but also represents a scalable solution. The system is suitable for a wide range of applications, from personal security to remote healthcare management, to the creation of smart homes. In the future, with the improvement of technologies and the adoption of faster, more accurate models, automatic activity analysis via Wi-Fi will become the standard, contributing to making daily life safer and more intelligently monitored without invasiveness.
Multivariate Time Series Analysis for Electrical Power Theft Detection in the Distribution Grid
A convolutional neural network analyzes multivariate time series to detect energy theft in distribution grids effectively.
A. Ceschini, A. Rosato, F. Succetti, R. Araneo, M. Panella
Leggi il paper

Multivariate Time Series Analysis for Electrical Power Theft Detection in the Distribution Grid

A. Ceschini, A. Rosato, F. Succetti, R. Araneo, M. Panella
How artificial intelligence transforms the fight against energy theft
The challenge of energy theft is one of the most critical issues for distribution network operators. In addition to causing significant economic losses, theft can compromise the quality of electricity supply, lead to blackouts, and slow down the transition to sustainable energy. With the introduction of smart grids and advanced meters, it is now possible to tackle this issue using cutting-edge artificial intelligence technologies. This innovative study proposes a system based on deep neural networks, such as CNNs, to detect energy theft through the analysis of multivariate time series, demonstrating how AI can revolutionize the automatic detection of anomalies.
A solution based on real-world data
The developed system leverages CNNs to analyze data collected in real-world scenarios, such as industrial sites with verified theft incidents. The data, acquired from multiple sensors, include variables such as cumulative active and reactive energy, total monthly power, and average power calculated across different time slots. These datasets, manually labeled based on certified reports, enable the model to identify anomalies in consumption and pinpoint the moments when theft occurs. Thanks to its multivariate structure, the system can uncover hidden correlations among variables, providing a more comprehensive view compared to traditional methods.
Promising performance in real-world conditions
The system was tested on five years of data, encompassing consumption profiles of users with 548 sequences sampled monthly. With an average accuracy of 76.1% in binary classification and 78.4% in a multiclass problem, the model has proven to be reliable and robust. Despite the complexity of the analyzed sequences and the variability of the case studies, the CNN effectively distinguished between regular and irregular consumption, significantly reducing error margins.
A future without waste
This technology is not limited to detecting theft; it marks a significant step toward more efficient and resilient management of distribution networks. By automating traditionally slow and costly processes, such as manual meter inspections, the system helps reduce economic and energy losses. Additionally, the scalability of the model makes it suitable for diverse contexts, from urban to industrial areas, seamlessly integrating with modern smart grid infrastructures.
Few-shot Federated Learning in Randomized Neural Networks via Hyperdimensional Computing
Fast, private AI learning from few examples using hyperdimensional computing and randomized networks across distributed devices.
A. Rosato, M. Panella, E. Osipov, D. Kleyko
Leggi il paper

Few-shot Federated Learning in Randomized Neural Networks via Hyperdimensional Computing

A. Rosato, M. Panella, E. Osipov, D. Kleyko
Learning Without Centralization: A Modern Necessity
In today’s hyperconnected world, the ability to train intelligent models without centralizing data is not just an advantage, it’s a necessity. Think of privacy, the communication limits of edge networks, or devices with limited computational power. This is where a new frontier is emerging: a distributed system powered by randomized neural networks and enhanced through hyperdimensional encoding, capable of learning quickly from just a few examples while sharing only the bare minimum.
Randomized Neural Networks Meet Hyperdimensional Encoding
At the heart of this system are models known as Random Vector Functional Links, which can be trained extremely fast and without backpropagation. But the true innovation lies in how these models share knowledge: instead of transmitting raw data or large model weights, they exchange a compressed, hyperdimensional representation of their classifier, produced through brain-inspired operations like binding and superposition. This leads to lower network traffic, increased privacy, and faster learning.
Few-Shot Learning
Hyperdimensional Computing doesn’t just compress efficiently, it performs surprisingly well. Tests across more than 100 real-world datasets showed that the system maintains an average accuracy above 70%, even in challenging scenarios with uneven data distribution across nodes. In certain network topologies, such as ring structures, it exceeds 80% accuracy in complex cases, thanks to its ability to merge local knowledge into a robust global model. This solution proves reliable even under extreme conditions: from networks of 10 to 100 nodes, the model remains stable and accurate, confirming its scalability. All this is achieved with minimal communication, often just two exchanges per node.
Endless Real-World Applications
Real-world applications? Practically endless. From healthcare to smart cities, environmental monitoring to wearables, anywhere that requires fast, distributed, privacy-respecting learning. A bold step toward lighter, more collaborative, and human-centered artificial intelligence.
06/12/2022
A randomized deep neural network for emotion recognition with landmarks detection
Novel randomized DNN uses facial landmarks for fast emotion recognition.
F. Di Luzio, A. Rosato, M. Panella
Leggi il paper
06/12/2022

A randomized deep neural network for emotion recognition with landmarks detection

F. Di Luzio, A. Rosato, M. Panella
The Importance of Emotions and AI's Role
In the complex tapestry of human interaction, emotions play a leading role. They are the compass guiding our decisions, coloring our perceptions, and shaping our relationships. Correctly recognizing and interpreting the moods of others is an intrinsically human ability, fundamental for empathy and effective communication. But what if machines could also develop such sensitivity? AI is opening previously unimaginable frontiers in this field.
Decoding Dynamic Facial Expressions
The human face is an incredibly rich expressive canvas. The subtlest muscle contractions, the tilt of the eyebrows, the curve of the lips: every detail can convey valuable information about a person's emotional state. AI is proving particularly adept at analyzing these micro-expressions, going beyond the simple interpretation of static images. Advanced technologies can now process video sequences, capturing the dynamics of emotional expression as it evolves over time.
The Facial Landmark Approach
An innovative approach relies on identifying and tracking hundreds of "landmarks" on the face. Imagine an incredibly detailed digital map of the face, composed of nearly 500 key points. By monitoring the precise coordinates of these points and their variation from one frame to the next, AI algorithms can reconstruct the emotional flow with remarkable accuracy, distinguishing between states like joy, sadness, surprise, anger, fear, or disgust. This method allows capturing not only the peak emotion but also the transition from a neutral state to the full expression.
Balancing Accuracy and Speed with Randomization
But how can these analyses be made not only accurate but also fast and efficient, especially for applications requiring real-time responses? This is where a fascinating concept comes into play: "randomized" neural networks. Instead of meticulously training every single parameter of the neural network, a process that can require significant time and computational resources, some parts of the network are set with random values and then "frozen." This approach, while potentially involving a very slight reduction in theoretical accuracy, drastically speeds up the learning phase and the algorithm's execution. It's a smart trade-off between accuracy and speed, crucial for bringing these technologies into the real world, onto devices with limited capabilities, or in scenarios where latency is critical.
Transformative Real-World Applications
The potential applications of AI capable of "reading" emotions are vast and transformative. Consider the healthcare sector: systems capable of monitoring subtle changes in facial expressions could aid in the early diagnosis of conditions like depression or chronic stress, supporting doctors and patients. Imagine e-learning systems that adapt educational materials based on the student's emotional reaction, making learning more personalized and effective. Or even more intuitive and empathetic human-machine interfaces, capable of responding more appropriately to our moods, enhancing user experience in countless contexts, from recommendation systems to road safety.
Towards Emotionally Aware Technology
The integration of AI into emotion recognition is not science fiction. By leveraging detailed facial landmark analysis and the efficiency of randomized networks, we are building systems increasingly capable of understanding the complex language of human emotions. This opens doors to a future where technology is not only functional but also more aware and attuned to our emotional experiences, promising significant improvements in crucial areas like health, education, and daily interaction.
Challenges and perspectives of smart grid systems in islands: a real case study
Integrating renewables with AI tools offers sustainable solutions, especially in isolated contexts.
F. Succetti, A. Rosato, R. Araneo, G. Di Lorenzo, M. Panella
Leggi il paper

Challenges and perspectives of smart grid systems in islands: a real case study

F. Succetti, A. Rosato, R. Araneo, G. Di Lorenzo, M. Panella
Renewable Energy Challenges for Island Systems
Islands like Ponza face unique challenges in achieving energy sustainability, primarily due to their reliance on imported diesel fuel. This dependency leads to high costs and frequent instability. Transitioning to renewable energy sources (RESs), such as photovoltaic (PV) and wind power, offers a promising alternative. However, intermittent energy production, limited storage options, and environmental constraints complicate the integration of RESs. Advanced technologies, including machine learning and deep learning, are instrumental in predicting energy demand and optimizing grid management.
The Role of Storage and Grid Innovations
Battery Energy Storage Systems (BESSs) play a pivotal role in stabilizing power grids with high RES penetration. For Ponza, the planned BESS infrastructure will provide a spinning reserve, manage fluctuations in RES production, and ensure consistent electricity supply during peak tourist seasons. Modernizing the grid further involves integrating automation, enabling real-time adjustments to power flows, and facilitating the seamless incorporation of new RES installations.
Predictive Insights Through AI
Artificial intelligence and deep learning techniques are applied to forecast energy production and demand accurately. These predictions are essential for designing efficient energy storage solutions and ensuring optimal use of RESs. On Ponza, machine learning models estimate future energy demand, incorporating variables like tourist influx, seasonal trends, and new electricity-driven initiatives such as electric vehicle charging stations and water desalination units.
Pathways to Sustainability
Achieving the 2030 targets outlined for Ponza involves significant upgrades to energy infrastructure. These include deploying RES systems with a capacity of up to 2.16 MW, installing advanced BESSs, and modernizing grid operations. By addressing environmental and logistical constraints, this approach aligns energy production with demand, reduces diesel reliance, and fosters long-term sustainability. The integration of smart grid systems ensures adaptability and efficiency, setting a benchmark for other islands globally.
Perceptron Theory Can Predict the Accuracy of Neural Networks
A perceptron-based theory predicts neural network accuracy using output statistics, fast, data-free, and surprisingly precise.
D. Kleyko, A. Rosato, E. Paxon Frady, M. Panella, F. T. Sommer
Leggi il paper

Perceptron Theory Can Predict the Accuracy of Neural Networks

D. Kleyko, A. Rosato, E. Paxon Frady, M. Panella, F. T. Sommer
From Black Boxes to Predictable Models
In recent years, deep neural networks have become the invisible engine behind much of today’s digital innovation, from image processing to autonomous driving. Yet, despite their effectiveness, they largely remain “black boxes”: powerful, but opaque. What if we could predict a model’s accuracy before fully training or testing it?
A Classic Theory Reimagined for Modern AI
This is where a revolutionary idea comes into play: reviving perceptron theory, one of the simplest structures in artificial intelligence, to predict the performance of complex neural networks. This modern statistical formulation, designed to operate on the output layer of any architecture, from recurrent networks to deep CNNs, allows for highly accurate estimates of a model's classification performance. The prediction relies only on a few statistical moments (such as mean and variance) of the postsynaptic sums, without needing to train any additional models.
High Accuracy Proven Across Models and Datasets
The results are impressive: this approach has been validated on over 120 real-world classification datasets and around 15 pretrained deep networks on ImageNet. In the latter case, the correlation between predicted and actual accuracy reaches 93%, with peaks of 97% using refined statistical approximations. Even in complex networks like ResNet, VGG, or NASNet, the method captures the model’s behavior with striking fidelity.
Scalable, Lightweight, and Data-Agnostic
Beyond precision, this model is both scalable and efficient: it can be applied to networks with thousands of output classes, reducing the entire evaluation process to a fast and non-invasive statistical analysis. In scenarios where data access is restricted (e.g., due to privacy), a variant of the method can estimate performance using only the final layer weights, still achieving notable predictive power.
A Step Toward More Transparent AI
The potential applications are wide-ranging: from selecting the best model in resource-constrained environments to building tools for explaining network decisions, and even detecting adversarial examples or outliers through the statistical behavior of activations.
Old Foundations for a New AI Era
In an era dominated by ever-larger and more complex models, rediscovering the simplicity and predictive power of a theory born over sixty years ago, and adapting it to modern AI, is not just an elegant intellectual move. It’s a decisive step toward more interpretable, trustworthy, and explainable neural networks.
02/08/2023
Modular quantum circuits for secure communication
Quantum modular circuits enable ultra-secure communication for fast, parallel encryption and decryption.
A. Ceschini, A. Rosato, M. Panella
Leggi il paper
02/08/2023

Modular quantum circuits for secure communication

A. Ceschini, A. Rosato, M. Panella
The Era of Quantum Cryptography
In the digital age, communication security is more crucial than ever. With the rise of cyber threats, traditional cryptography methods are becoming obsolete. The advent of Quantum Computing offers new prospects, enabling the creation of more advanced and impenetrable security systems. Among these innovations, quantum quasi-chaotic (QC) generators play a central role, capable of producing pseudorandom sequences with extremely high entropy, ideal for encrypting information in a highly secure manner.
Quantum Modular Circuits: The Key to Unbreakable Security
Current technologies rely on encryption systems that, although sophisticated, remain vulnerable to advances in quantum computing. Quantum modular circuits exploit modular arithmetic in a quantum context to generate and manage cryptographic keys in parallel, increasing the speed and robustness of encryption. At the heart of this approach is a nonlinear digital filter implemented through quantum modular operations that allow for the creation of a pseudorandom behavior ideal for encrypting and decrypting information. By using Quantum Modular Addition and Multiplication, these circuits can perform cryptographic transformations with inherent security far beyond that of classical methods, leveraging quantum superposition and entanglement to process multiple data streams simultaneously. This enables the construction of multi-channel encryption and decryption systems capable of handling several data flows in parallel without compromising performance.
Real-World Applications: Ultra-Secure Communications and Quantum Cryptography
The applications of these systems range from protecting government telecommunications to securing banking and financial networks. In the military sector, quantum quasi-chaotic generators could ensure secret transmissions that are impossible to decipher using conventional technologies. In telecommunications, their implementation could revolutionize end-to-end encryption, safeguarding sensitive data transmission even in highly vulnerable environments. Another key area is the Internet of Things (IoT), where billions of connected devices require advanced protection against intrusions. Quantum cryptography based on modular circuits represents a scalable solution that can ensure security even in the most complex and distributed systems.
Towards a Future of Quantum Communications
The large-scale implementation of these circuits is still in development, but experimental results show that quantum QC generators can outperform traditional solutions in terms of security and efficiency. With the advancement of quantum hardware and the development of error mitigation techniques, these systems could form the foundation of future ultra-secure communication networks. The adoption of modular quantum circuits for cryptography is not just a technological improvement, but a true revolution in digital security. In an increasingly connected world vulnerable to cyberattacks, quantum technologies represent the only path to ensuring absolute data protection.
An adaptive embedding procedure for time series forecasting with deep neural networks
A novel deep learning model that integrates adaptive embedding with bidirectional LSTMs to enhance time series forecasting.
F. Succetti, A. Rosato, M. Panella
Leggi il paper

An adaptive embedding procedure for time series forecasting with deep neural networks

F. Succetti, A. Rosato, M. Panella
More Reliable Forecasting with AI
From finance to energy and meteorology, time series forecasting is a crucial challenge across various industries. Analyzing historical data to predict future trends is often hindered by non-linearity, high variability, and long-term dependencies. Deep neural networks have proven to be powerful tools for tackling these issues, but their effectiveness is often limited by inefficient data handling and a heavy reliance on parameter optimization. A new approach based on adaptive embedding is revolutionizing time series forecasting by enhancing both accuracy and flexibility. This technique automatically extracts a compressed representation of historical data, reducing problem complexity and optimizing the predictive capabilities of neural networks.
LSTM and Adaptive Embedding: How It Works
At the core of this innovation lies a bidirectional Long Short-Term Memory (LSTM) network, structured in two layers. The first layer performs adaptive embedding, identifying the most relevant patterns in the time series without human intervention. This pre-training phase enables the model to better understand the underlying data structure. The second layer then uses this information to make more precise predictions. The key idea is to eliminate the need for separate feature extraction algorithms, integrating data analysis directly within the neural network. This not only simplifies the process but also makes the system more efficient and applicable to any context, from financial market fluctuations to intelligent energy management.
Real-World Applications
This approach has been tested in real-world scenarios, demonstrating high accuracy and versatility. It has been successfully applied to forecasting energy consumption, photovoltaic production, and financial data. The results show that the model reduces forecasting errors compared to traditional techniques, significantly improving its ability to adapt to highly dynamic data. A crucial advantage is its generalization capability: the system can be used across different industries without requiring extensive customization. This makes it a powerful tool for businesses and institutions that need reliable forecasts to optimize resource management.
Towards More Efficient Predictive Intelligence
The integration of adaptive embedding and deep neural networks marks a breakthrough in the world of forecasting. This approach not only improves model accuracy but also reduces computational load, making it ideal for real-time applications. With this technology, AI is becoming an increasingly strategic tool for understanding and anticipating complex phenomena, helping businesses and researchers make more informed, data-driven decisions.
08/08/2024
A Neural Network Symbolic Approach to Structural Health Monitoring in Aerospace Applications
A symbolic deep learning approach enhances structural health monitoring in aerospace achieving near-perfect damage classification.
F. Angeletti, F. Succetti, M. Panella, A. Rosato
Leggi il paper
08/08/2024

A Neural Network Symbolic Approach to Structural Health Monitoring in Aerospace Applications

F. Angeletti, F. Succetti, M. Panella, A. Rosato
Structural Monitoring: A Challenge for Space Missions
Ensuring the structural integrity of spacecraft and satellites is one of the most critical challenges for aerospace missions. Traditional visual inspection methods are often impractical in orbit, making it necessary to adopt advanced solutions based on distributed sensors and automated analysis algorithms. Among the main threats to the stability of space structures are impacts with orbital debris, extreme thermal oscillations, and material fatigue, all of which can compromise the proper functioning of satellites and their appendages, such as solar panels.
AI and Neural Networks for Damage Detection
The integration of Artificial Intelligence into structural health monitoring represents a paradigm shift. The use of deep neural networks, particularly Long Short-Term Memory (LSTM) networks, enables the analysis of time-series data from accelerometers and other sensors, identifying anomalies that could indicate structural damage. The adopted model leverages an innovative approach, combining symbolic representation of time series with a recurrent neural network architecture. By compressing information through Symbolic Aggregate approXimation (SAX), data dimensionality is reduced, improving processing speed and enhancing the model’s ability to recognize recurring patterns. This method transforms complex time-series data into symbolic sequences, simplifying the classification process and making the algorithm more robust to data variations.
Applications in Space: From Simulation to Reality
To test the system’s effectiveness, a simulation of a satellite with flexible solar panels equipped with acceleration sensors was conducted. The model was evaluated under multiple damage scenarios, replicating realistic conditions of impacts with space debris. The results showed that using symbolic representation enhances the accuracy of damage classification, achieving an almost 100% precision and significantly reducing false positives and negatives. The ability of an AI model to autonomously and reliably detect structural damage without human intervention is a crucial step forward for space missions. The combination of recurrent neural networks and dimensionality reduction techniques opens new possibilities for onboard automatic monitoring, essential for ensuring the safety and longevity of space infrastructures.
Towards a Future of Intelligent Monitoring
The adoption of AI in the aerospace sector is transforming the way space systems are monitored and managed. The ability to implement autonomous solutions based on deep learning and symbolic representation reduces computational load while improving diagnostic reliability. With advancements in onboard computing technologies, these solutions could become standard for satellite and space module monitoring, paving the way for a new era of autonomous and intelligent space exploration.
A variational approach to quantum gated recurrent units
A faster and efficient Quantum Gated Recurrent Unit (QGRU) improves time series forecasting.
A. Ceschini, A. Rosato, M. Panella
Leggi il paper

A variational approach to quantum gated recurrent units

A. Ceschini, A. Rosato, M. Panella
Quantum Artificial Intelligence Revolutionizing Predictions
From finance to renewable energy, time series forecasting is a cornerstone for optimizing strategic decisions and improving resource management. However, traditional deep learning models, such as Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks, face significant limitations: high computational costs, long training times, and challenges in handling long-term dependencies. The integration of Quantum Artificial Intelligence (QAI) with RNNs opens new possibilities, leveraging quantum superposition and entanglement to enhance computational efficiency and forecasting accuracy. An innovative architecture based on Quantum Gated Recurrent Units (QGRU) introduces a faster and more efficient model compared to both classical and existing quantum alternatives.
Quantum GRU: A Faster and More Efficient Model
The QGRU architecture is based on Variational Quantum Circuits (VQC), which process temporal data in a high-dimensional space, maximizing the potential of current quantum devices. The model combines parametric quantum layers with two classical preprocessing and postprocessing layers, optimizing data input and output. One of the main advantages of this architecture is the 25% reduction in quantum parameters compared to Quantum LSTM (QLSTM) networks. This translates into greater computational efficiency: the QGRU model is about 25% faster in both training and inference compared to QLSTM, making it more suitable for implementation on real quantum hardware and simulators.
Real-World Applications: From Meteorology to Energy
The effectiveness of QGRU has been tested across various real-world scenarios, demonstrating its superiority over both classical models and other quantum solutions. One of the most challenging applications is the prediction of solar cycles, where sunspots exhibit high variability and make forecasting particularly complex. The quantum model has shown a remarkable ability to adapt to these fluctuations, outperforming classical neural networks in handling noisy and nonlinear data. In the field of renewable energy, QGRU has been applied to wind power generation, an area where rapid and unpredictable variations pose significant challenges for grid management. The model has proven capable of producing more stable and reliable predictions, reducing the average forecasting error by 40% compared to conventional LSTM networks. This improved accuracy is crucial for optimizing energy distribution and integrating renewables more efficiently into the electrical grid. QGRU has also been tested on periodic time series, demonstrating superior stability and modeling capabilities. Unlike traditional approaches, which struggle with long-term dependencies and complex temporal patterns, the quantum model effectively captures underlying trends, offering a more robust and adaptable solution for time series forecasting in multiple domains.
Towards a Quantum Future for Deep Learning
The implementation of quantum recurrent neural networks marks a significant step forward in time series forecasting, offering a winning combination of accuracy and computational speed. As quantum hardware continues to evolve, these architectures could become increasingly accessible, paving the way for new applications in strategic sectors such as finance, healthcare, and energy. QAI is transforming how we interpret data and make decisions, ushering machine learning into a new era. The future of forecasting? Faster, more precise, and… quantum-powered.
Quantum enhanced knowledge distillation
Classical-to-quantum knowledge distillation boosts hybrid AI performance using efficient quantum circuits and reduced model sizes.
S. Piperno, L. Lavagna, F. De Falco, A. Ceschini, A. Rosato, D. Windridge, M. Panella
Leggi il paper

Quantum enhanced knowledge distillation

S. Piperno, L. Lavagna, F. De Falco, A. Ceschini, A. Rosato, D. Windridge, M. Panella
Bridging the Classical and Quantum Worlds
In a context where quantum computing still faces the limitations of Noisy Intermediate-Scale Quantum (NISQ) hardware, the challenge is clear: how can quantum Machine Learning be made truly useful today? The answer comes from a technique already well-known in the classical AI world, but rarely explored at scale in the quantum domain: Knowledge Distillation (KD).
Teaching Through Structure, Not Just Labels
In essence, KD is a process where a powerful and complex model (the “teacher”) guides the training of a simpler model (the “student”), transferring not just labels but structured insights learned during classification. The breakthrough is that this knowledge transfer can now take place from a classical model (like an MLP) to a hybrid quantum-classical model, effectively bridging two previously distant worlds.
Smaller Models, Smarter Results
The case study focuses on a non-linearly separable multi-class classification problem built on an extended XOR dataset. The teacher model is an MLP with over 1200 parameters, while the student can either be a simplified classical neural network or a hybrid structure with variational quantum circuits. Thanks to KD, the student model learns more efficiently: performance improves significantly, even in complex scenarios, with a substantial reduction in parameter count (from 195 in the classical version to just 74 in some quantum architectures).
Quantum Students That Learn Better
Performance metrics like the average F1 score confirm the value of the approach: even the lightest quantum models, typically penalized due to limited parameters and sensitivity to initialization, clearly benefit from distillation. The tested configurations, from universal circuits to setups with selective measurement-based compression, prove that this method works and paves the way for more effective quantum systems, even on constrained hardware.
A Strategic Vision for Hybrid AI
This form of “cross-domain teaching” is not just a pragmatic solution for the present; it’s a strategic vision for the future. Until quantum computers reach full maturity, classical AI can serve as a guide. And when knowledge transfer is optimised through compatible structures, hybrid models become competitive, sustainable, and far more intelligent.
14/11/2024
An explainable fast deep neural network for emotion recognition
A fast, explainable deep neural network enhances emotion recognition by optimizing facial landmark analysis.
F. Di Luzio, A. Rosato, M. Panella
Leggi il paper
14/11/2024

An explainable fast deep neural network for emotion recognition

F. Di Luzio, A. Rosato, M. Panella
Facial Expressions: The Key to Non-Verbal Communication
Facial expressions are the foundation of human non-verbal communication. A smile, a look of surprise, an expression of disgust—each emotion is conveyed through subtle facial micro-movements that, until recently, were difficult for machines to interpret with precision. Today, thanks to an innovative deep neural network, not only can emotions be recognized with high accuracy, but this can now be done quickly and, most importantly, in an explainable manner.
From Opaque Intelligence to Understandable Intelligence
Deep neural networks have always faced a major challenge: their "black box" nature. While they make highly accurate predictions, their decision-making processes have long remained obscure. This new model overcomes this barrier through the use of explainable AI, which can identify and prioritize the most relevant features for emotion recognition.By integrating Integrated Gradients, an advanced Explainable AI technique, the model can analyze the contribution of each facial reference point (landmark) in classifying emotions. The original input consists of 468 facial landmarks extracted from video sequences, but through data relevance analysis, the system can reduce the number of utilized points without compromising accuracy. Tests have shown that by reducing the landmarks to 128, the model maintains an accuracy above 97% for certain emotions, such as surprise and happiness, while significantly lowering computational costs.
Computational Efficiency and Performance
Optimizing the number of features is not just a theoretical exercise; it leads to tangible improvements in model performance. Tests conducted on standard datasets such as CK+ have revealed a significant reduction in inference time, making the system suitable for real-time applications.
Real-World Applications: From Healthcare to Security
The applications of this technology go far beyond basic facial recognition. In the medical field, a more precise analysis of facial expressions can support neurological and psychiatric diagnoses by detecting early signs of emotional disorders. In security systems, real-time monitoring of facial expressions can help identify suspicious intent, contributing to safer public spaces.The optimization of AI models through explainability techniques represents a major leap forward: not only does it ensure more reliable and transparent functioning, but it also paves the way for a future where AI is not just intelligent but also comprehensible and accessible.
A Deep Learning-based Approach for Battery Life Classification
A deep learning-based LSTM network accurately classifies battery health, optimizing energy storage and predictive maintenance.
F. Succetti, A. Dell'Era, A. Rosato, A. Fioravanti, R. Araneo, M. Panella
Leggi il paper

A Deep Learning-based Approach for Battery Life Classification

F. Succetti, A. Dell'Era, A. Rosato, A. Fioravanti, R. Araneo, M. Panella
Artificial Intelligence for Energy Management
Batteries are the core of modern energy infrastructures, from storage systems for renewables to electric vehicles. Monitoring their health and predicting degradation is essential to ensure efficiency, safety, and operational longevity. However, the complex and nonlinear nature of batteries makes accurate diagnosis difficult with traditional methods. Artificial Intelligence offers an innovative solution by leveraging deep neural networks to analyze charge and discharge cycles and classify battery health with high precision.
LSTM and Deep Learning for Health State Classification
The use of Long Short-Term Memory (LSTM) networks allows for modeling the temporal dynamics of batteries, capturing long-term relationships between electrical parameters such as voltage and current. This architecture, optimized for time series analysis, enables the classification of battery degradation levels, from "new" to "old," facilitating predictive maintenance and energy storage management. The approach utilizes real laboratory data to train the model, relying on charge and discharge cycles recorded in a controlled environment. Data preprocessing, including time alignment and normalization, ensures accurate analysis and reduces the model’s sensitivity to variations in input data.
Real-World Applications: From Industry to Electric Vehicles
The ability to accurately predict battery health has a direct impact on numerous sectors. In the renewable energy field, it optimizes energy storage usage, preventing overloads and improving grid management. In automotive applications, it enables better battery management in electric vehicles, increasing range and reducing replacement costs. In industrial settings, it helps prevent failures in battery-powered devices, enhancing equipment reliability.
Towards AI-Driven Predictive Maintenance
The integration of AI into battery management marks a significant step toward more autonomous and efficient systems. With increasing data availability and increasingly accurate models, it will be possible to develop predictive solutions that dynamically adapt to real-world operating conditions. Deep learning-based diagnostics represent a key innovation for the future of energy, ensuring greater sustainability and reliability in storage systems.
05/12/2024
Enhancing Autism Detection Through Gaze Analysis Using Eye Tracking Sensors and Data Attribution with Distillation in Deep Neural Networks
A deep learning model enhances early autism diagnosis by analyzing visual patterns with eye tracking.
F. Colonnese, F. Di Luzio, A. Rosato, M. Panella
Leggi il paper
05/12/2024

Enhancing Autism Detection Through Gaze Analysis Using Eye Tracking Sensors and Data Attribution with Distillation in Deep Neural Networks

F. Colonnese, F. Di Luzio, A. Rosato, M. Panella
Beyond Traditional Diagnosis: The Role of AI in Autism Detection
Autism Spectrum Disorder (ASD) is typically diagnosed through behavioral assessments, structured questionnaires, and clinician observations. While effective, these methods rely heavily on subjective interpretation, which can lead to variability in diagnostic outcomes. The integration of artificial intelligence into medical diagnostics opens up new possibilities for more precise and objective screening, especially when combined with advanced technologies such as eye tracking.
Decoding Gaze Patterns with AI
Individuals with ASD often exhibit unique gaze behaviors, such as reduced eye contact or a preference for focusing on peripheral objects rather than social stimuli like faces. Eye tracking sensors provide a powerful tool to quantify these behaviors, capturing precise information about how a person scans and fixates on different elements in a visual scene. By leveraging deep neural networks, these gaze patterns can be analyzed in real-time, identifying characteristics that distinguish individuals with ASD from neurotypical individuals.
Optimizing Accuracy Through Data Selection
Processing large datasets for AI training can be computationally expensive and time-intensive. To enhance efficiency, an innovative approach known as data attribution is used, allowing AI models to prioritize the most relevant training samples while filtering out noisy or misleading data. By applying a technique called TracIn, researchers can evaluate how each data point influences the model’s learning process, refining the dataset without compromising accuracy. In fact, results show that even when trained on just 77% of the dataset, the model maintained a classification accuracy of 94.35%, surpassing benchmarks and proving that selecting high-quality data is more effective than simply increasing the dataset size.
From Lab to Real-World Applications
This technology has the potential to transform autism screening and diagnosis. AI-powered gaze analysis could be implemented in clinical settings, providing clinicians with an additional, objective tool to support early detection. It could also be integrated into portable diagnostic devices, making autism screening more accessible in schools or pediatric clinics. Moreover, by identifying the most influential gaze patterns linked to ASD, this research enhances the broader understanding of visual attention differences, contributing to improved therapeutic approaches.
A Future of AI-Assisted Diagnosis
The combination of AI, deep learning, and eye tracking represents a major step toward more reliable and interpretable medical AI applications. By improving accuracy while reducing computational overhead, this approach not only refines ASD classification but also lays the groundwork for integrating AI-driven insights into clinical practice. In the near future, these models could be adapted for other neurodevelopmental conditions, further bridging the gap between AI innovation and healthcare.

Vuoi saperne di più?

Entra in contatto con noi e descrivici le tue esigenze. Saremo lieti di chiarire ogni tuo dubbio e aiutarti ad accellerare la crescita del tuo business.
Contattaci