Hyperdimensional Computing (HDC) is a neuro-inspired computational model that represents and manipulates information using high-dimensional distributed representations that are combined and compared using simple and highly parallel vector operations. In this work, we present a highly flexible hardware acceleration unit for HDC learning tasks based on the binary spatter-code model. Integrated into the execution stage of the Klessydra-T03 RISC-V core, the unit accelerates the core arithmetic operations of HDC and can be configured at synthesis time in terms of hardware parallelism, supported operations, and size of the local memories, trading off execution time with hardware resources to match application needs. A custom RISC-V Instruction Set Extension efficiently controls the accelerator, with instructions fully integrated into the GNU Compiler Collection toolchain and exposed to the programmer as intrinsics. Dedicated Control and Status Registers allow specifying the characteristics of the high-dimensional space and the target learning tasks at runtime, controlling the hardware loops of the accelerator and enabling the same hardware architecture to be used for various tasks. The dual flexibility coming from hardware configuration and software programmability sets this work apart from application-specific solutions in the literature, offering a unique, versatile accelerator adaptable to a wide range of applications and learning tasks.

As the Internet of Things expands, the need to embed artificial intelligence algorithms into resource-constrained devices for real-time applications is growing. These systems require efficient and scalable algorithms to perform rapid and autonomous decision-making without relying on centralized cloud servers. Hyperdimensional Computing (HDC) has recently emerged as a compelling paradigm for learning tasks in such environments, offering computational efficiency, exceptional parallelism, and scalability.In this work, we present HD-CBBIN, a lightweight and efficient implementation of the HD-CB framework for modeling and automating sequential decision-making Contextual Bandits (CB) problems on embedded systems. By introducing modifications to the original algorithm, HD-CBBINexclusively uses binary hypervectors, significantly reducing computational demands. We benchmark the performance of HD-CBBINon synthetic datasets, comparing it to the real-valued counterpart and the traditional state-of-the-art LinUCB algorithm. We also evaluate its execution time, computational complexity, and memory requirements on various embedded platforms and demonstrate additional gains through hardware acceleration. The results show that our approach achieves linear execution time with respect to the context vector size and up to a 141× speedup over LinUCB while maintaining competitive performance, establishing a new milestone in contextual bandit algorithms for time-critical, resource-constrained applications.

In this work, we introduce a hybrid quantum-classical framework to address anomaly detection problems in time series data using an innovative Quadratic Unconstrained Binary Optimization formulation. The proposed approach integrates density-based and statistical methodologies with the Quantum Approximate Optimization Algorithm, providing a versatile tool for anomaly detection. Moreover, the underlying model is transversal to different kinds of anomalies and can be directly applied to diverse applications, ranging from fault detection to novelty discovery. Experimental results prove that this architecture achieves competitive accuracy compared to classical techniques, with the unique advantage of exploring alternative solutions in parallel due to the properties of quantum computing. This capability enables a deeper understanding of patterns in time series data.

Gait analysis is an important technique for diagnosing, monitoring and rehabilitating neurological conditions such as Autism Spectrum Disorder, also known as ASD. With the increasing employment of AI in the medical domain, gait analysis allows researchers and doctors to improve their ability in modeling and interpreting complex structured data. In this delicate domain, explainable models become essential to build medical frameworks that can fasten and objectify the diagnosis procedure which is generally slow and ineffective. This study explores the use of explainability techniques applied to graph neural networks to enhance the understanding of decisions made by gait analysis models to detect ASD. The findings demonstrate that integrating explainability AI tools in this particular domain increases the accuracy of ASD detection and improves model transparency of the entire diagnosis process, allowing specialists to interpret and validate the extracted information and facilitating the adoption of these models in clinical settings.

In this work, we introduce new quantum machine learning models that combine both quantum and hyperdimensional computing. We focus our effort on two novel architectures that are first theoretically demonstrated, and then applied for testing to prototypical machine learning tasks, namely for pattern completion, classification, and clustering. We obtained accurate and promising results that prove, for the first time, the synergies between two of the most innovative computational approaches such as quantum computing and hyperdimensional computing.

Knowledge distillation is a widely explored technique in classical machine learning, in which a smaller or more efficient model is trained to mimic the behavior of a larger, more complex model. In this study, we extend the concept of knowledge distillation from classical architectures to quantum architectures, with the goal of improving the training of quantum models while potentially reducing the number of parameters compared to their classical counterparts. Given the inherent challenges in training quantum neural networks, leveraging knowledge from well-established classical models could provide valuable insights and advantages, particularly in terms of model efficiency and performance. In this work we explore the potential benefits of this approach, evaluating a hybrid quantum model against a non-hybrid quantum baseline. While the proposed study is still in the preliminary stage, it aims to set the scene for further investigation into the most appropriate architecture for classical-to-quantum knowledge distillation in order to enhance the development and optimization of quantum neural networks more generally.

This paper investigates the potential of quantum-enhanced models for time series forecasting in environments where time is a fundamental constraint. The focus is on on the quantum long short term memory networks with reservoir (QR-LSTM), and our proposed novel extension, the Autoencoded Quantum Reservoir LSTM Model (QR-LSTM-AE). The QR-LSTM, while offering substantial computational benefits in terms of time and resource efficiency, incurs a slight decrease in performance relative to the fully trainable quantum LSTM model. To address this trade-off, we introduce the QR-LSTM-AE, which utilizes an autoencoder-based strategy to recover performance lost in the reservoir alone, achieving superior accuracy without the need for full retraining. This approach not only maintains the efficiency of the QR-LSTM but also significantly reduces the computational cost associated with adapting to new data or time series, making it ideal for real-time forecasting applications. The experiments are carried out on a energy-related, highly volatile dataset, and results underscore the importance of balancing predictive accuracy with computational efficiency, highlighting the potential of quantum models to offer practical solutions in time-constrained environments. Our findings demonstrate that the QR-LSTM-AE effectively balances predictive accuracy and computational efficiency, paving the way for future advancements in quantum-enhanced forecasting through self-adaptive models, quantum autoencoders, and attention-based reservoir computing.

Early detection and accurate diagnosis of Autism Spectrum Disorder are critical for enabling timely interventions and personalized support. However, conventional diagnostic approaches remain largely subjective, posing challenges in achieving consistency and reliability. Recent clinical studies suggest that eye gaze patterns offer valuable insights into cognitive and attentional mechanisms underlying Autism Spectrum Disorder, highlighting their potential as biomarkers for objective assessment. In this work, we introduce a patch-based Graph Neural Network framework designed to analyze eye-tracking data by integrating fixation maps with the corresponding visual stimuli observed by both neurotypical and neurodivergent individuals. The proposed method constructs a hierarchical graph representation of image patches, progressively aggregating visual and attentional information to enhance feature extraction. A combination of Graph and Convolutional Neural Networks are employed for classification, capturing spatial and contextual dependencies between regions of interest respectively. Experimental evaluations and analysis demonstrate the effectiveness of our approach in distinguishing Autism Spectrum Disorder from neurotypical behavior, underscoring the potential of graph-based learning for neurodevelopmental assessments.

This paper explores the integration of quantum tools into classical cryptographic systems to assess their impact on the feasibility-efficiency trade-off in the context of Noisy Intermediate-Scale Quantum technology. By focusing on the hybridization of classical low-level primitives and high level cryptographic tools, the study investigates the effect of quantum contributions on encryption schemes corresponding to trapdoor permutations, revealing significant insights into the intersection of cryptography, quantum computing, and Boolean circuit theory.

Deep Learning models, and specifically Recurrent Neural Networks, have been successfully applied to time series classification in many applications, including Structural Health Monitoring. A relatively new field of research is the implementation of Deep Learning techniques for damage identification via measured time series in space systems, which proves chal-lenging in case of Structural Health Monitoring of large flexible structures. In this work, we propose a novel approach exploiting a symbolic time series representation as an additional data pre-processing step for data dimensionality reduction. The strategy is applied to a real world-scenario of a spacecraft hosting large solar panels equipped with distributed accelerometers at structural level, and compared against a previous benchmark case developed by the authors. Obtained results prove that the strategy has the potential to further improve classification quality towards an ideal 100% accuracy envisioned for space systems.

Hyperdimensional Computing (HDC) is a bio-inspired learning paradigm, that models neural pattern activities using high-dimensional distributed representations. HDC leverages parallel and simple vector arithmetic operations to combine and compare different concepts, enabling cognitive and reasoning tasks. The computational efficiency and parallelism of this approach make it particularly suited for hardware implementations, especially as a lightweight, energy-efficient solution for performing learning tasks on resource-constrained edge devices. The HDC pipeline, including encoding, training, and comparison stages, has been extensively explored with various approaches in the literature. However, while these techniques are mainly oriented to improve the model accuracy, their influence on hardware parameters remains largely unexplored. This work presents AeneasHDC, an automatic and open-source platform for the streamlined deployment of HDC models in both software and hardware for classification, regression and clustering tasks. AeneasHDC supports an extensive range of techniques commonly adopted in literature, automates the design of flexible hardware accelerators for HDC, and empowers users to easily assess the impact of different design choices on model accuracy, memory usage, execution time, power consumption, and area requirements.

This paper introduces a novel variational quantum algorithm built upon the established Quantum Approximate Optimization Algorithm also known as QAOA. Since the known parameter fixing strategy imposes constraints on QAOA to enhance tractability at the cost of some expressive power, the proposed layerwise approach integrates it with the existing Multi-Angle QAOA technique, which is characterized in turn by height-ened expressiveness through an increased number of parameters, albeit with increased optimization challenges. Consequently, the proposed layerwise-Multi-Angle QAOA combines the strengths of one variant with the limitations of the other, striking a balance in algorithmic design. The effectiveness of the proposed algorithm is assessed through experimental evaluations on the Maximum Cut problem. This study reveals promising results in heuristic predictions, with robustness both in terms of approximation ratio and optimization capabilities.

Autism Spectrum Disorder (ASD) is a complex and heterogeneous neurodevelopmental disorder which affects a significant proportion of the population, with estimates suggesting that about 1 in 100 children worldwide are affected by ASD. This study introduces a new Deep Neural Network for identifying ASD in children through gait analysis, using features extracted from frames composing video recordings of their walking patterns. The innovative method presented herein is based on imagery and combines gait analysis and deep learning, offering a noninvasive and objective assessment of neurodevelopmental disorders while delivering high accuracy in ASD detection. Our model proposes a bimodal approach based on the concatenation of two distinct Convolutional Neural Networks processing two feature sets extracted from the same videos. The features obtained from the convolutions of both networks are subsequently flattened and merged into a single vector, serving as input for the fully connected layers in the binary classification process. This approach demonstrates the potential for effective ASD detection in children through the combination of gait analysis and deep learning techniques.

Multilayer neural networks set the current state of the art for many technical classification problems. But, these networks are still, essentially, black boxes in terms of analyzing them and predicting their performance. Here, we develop a statistical theory for the one-layer perceptron and show that it can predict performances of a surprisingly large variety of neural networks with different architectures. A general theory of classification with perceptrons is developed by generalizing an existing theory for analyzing reservoir computing models and connectionist models for symbolic reasoning known as vector symbolic architectures. Our statistical theory offers three formulas leveraging the signal statistics with increasing detail. The formulas are analytically intractable, but can be evaluated numerically. The description level that captures maximum details requires stochastic sampling methods. Depending on the network model, the simpler formulas already yield high prediction accuracy. The quality of the theory predictions is assessed in three experimental settings, a memorization task for echo state networks (ESNs) from reservoir computing literature, a collection of classification datasets for shallow randomly connected networks, and the ImageNet dataset for deep convolutional neural networks. We find that the second description level of the perceptron theory can predict the performance of types of ESNs, which could not be described previously. Furthermore, the theory can predict deep multilayer neural networks by being applied to their output layer. While other methods for prediction of neural networks performance commonly require to train an estimator model, the proposed theory requires only the first two moments of the distribution of the postsynaptic sums in the output neurons. Moreover, the perceptron theory compares favorably to other methods that do not rely on training an estimator model.

Nowadays, solving time series prediction problems is an open and challenging task. Many solutions are based on the implementation of deep neural architectures, which are able to analyze the structure of the time series and to carry out the prediction. In this work, we present a novel deep learning scheme based on an adaptive embedding mechanism. The latter is exploited to extract a compressed representation of the input time series that is used for the subsequent forecasting. The proposed model is based on a two-layer bidirectional Long Short-Term Memory network, where the first layer performs the adaptive embedding and the second layer acts as a predictor. The performances of the proposed forecasting scheme are compared with several models in two different scenarios, considering both well-known time series and real-life application cases. The experimental results show the accuracy and the flexibility of the proposed approach, which can be used as a prediction tool for any actual application.

This paper aims at solving time series prediction problems by means of a hybrid quantum-classical recurrent neural network. We propose a novel architecture based on stacked Long Short-Term Memory layers and a variational quantum layer. The latter employs a quantum feature map to embed input data into quantum states, which are then processed by a circuit ansatz. Finally, the expectation value of the circuit's outcome is taken over Pauli observables. Quantum properties such as superposition and entanglement are exploited to perform computations efficiently in a high-dimensional feature space. The proposed hybrid quantum-classical neural network is applied to a real-life challenging problem pertaining to the prediction of renewable energy time series. The comparison between the proposed approach and the classical counterpart shows that the former achieves better results in terms of prediction error, thus demonstrating better approximation of stochastic fluctuations and an overall effectiveness of the quantum variational approach also for prediction tasks.

The recent interest in federated learning has initiated the investigation for efficient models deployable in scenarios with strict communication and computational constraints. Furthermore, the inherent privacy concerns in decentralized and federated learning call for efficient distribution of information in a network of interconnected agents. Therefore, we propose a novel distributed classification solution that is based on shallow randomized networks equipped with a compression mechanism that is used for sharing the local model in the federated context. We make extensive use of hyperdimensional computing both in the local network model and in the compressed communication protocol, which is enabled by the binding and the superposition operations. Accuracy, precision, and stability of our proposed approach are demonstrated on a collection of datasets with several network topologies and for different data partitioning schemes.

Nowadays, deep learning is gaining attraction as one of the most successful paradigm for a plethora of machine learning applications. While its benefits are undoubted, the high computational burden associated with its training algorithms and cross-validation procedures is stimulating new lines of research. To this end, randomized deep neural networks are one of the best alternatives in terms of efficiency-to-accuracy balance. In this paper, we present a deep neural architecture that uses randomization of some parameters in a complex structure whose novelty is twofold: it embeds past samples of the time series by using daily blocks in the input frame of a convolutional layer; it predicts a day on the whole by solving a suitable regression problem. The proposed randomized approach is compared with state-of-the-art prediction algorithms on the challenging context related to energy time series, where the day-ahead prediction is usual, obtaining comparable or even better results in terms of forecasting accuracy and training time.

In the supervised learning domain, considering the recent prevalence of algorithms with high computational cost, the attention is steering towards simpler, lighter, and less computationally extensive training and inference approaches. In particular, randomized algorithms are currently having a resurgence, given their generalized elementary approach. By using randomized neural networks, we study distributed classification, which can be employed in situations were data cannot be stored at a central location nor shared. We propose a more efficient solution for distributed classification by making use of a lossy compression approach applied when sharing the local classifiers with other agents. This approach originates from the framework of hyperdimensional computing, and is adapted herein. The results of experiments on a collection of datasets demonstrate that the proposed approach has usually higher accuracy than local classifiers and getting close to the benchmark - the centralized classifier. This work can be considered as the first step towards analyzing the variegated horizon of distributed randomized neural networks.

The need of reliable prediction algorithms of energy production is increasing due to the spread of smart solution for grid, plant and resource management. Recurrent neural networks are a viable solution for prediction but their performance is somewhat insufficient when the time series is generated by an underlying process that behaves in a complex manner. In this paper, a new combination of echo state network and genetic algorithms is employed in order to improve the prediction accuracy of photovoltaic time series. The genetic algorithm is used to embed past samples of the time series to be used for predicting a new one. It aims at a feature extraction in order to regularize data being fed into a neural network model, so that it is able to learn more robust and generalizable prediction models. The experimental tests prove that the proposed approach is suited to the application focused in this paper.

In this paper, we introduce Random Weights Fuzzy Neural Networks as a suitable tool for solving prediction problems. The generalization capability of these randomized fuzzy neural networks is exploited in order to estimate accurately the sample be predicted from a multidimensional input. The latter is obtained by applying an embedding technique to the time series, which selects only the meaningful past samples to be used for prediction. We tested the proposed approach on real-world time series pertaining to the application context of power delivery. We proved the efficacy of the proposed approach by comparing its forecasting accuracy with respect to other prediction systems based on well-known data-driven regression models.

In modern real-world applications, the need of using a decentralized data processing approach has progressively increased, facing complexity and handling issues. Pervasive data and ubiquitous computational capacity have enabled the proficient use of distributed implementation of machine learning algorithms, especially for forecasting problems. We provide in this paper a new, fully distributed prediction approach based on the Long Short-Term Memory deep neural network. When placed in a network of interconnected agents, the single predictors are able to improve the prediction accuracy by means of the Alternating Direction Method of Multipliers consensus procedure on some network parameters. Experimental tests on real-world time series prove the efficacy of the proposed approach, which regulates the information exchange in the network through high-level structures in the considered models.

Quantum computing ideally allows designers to build much more efficient computers than the existing classical ones. By exploiting quantum parallelism and entanglement, it is possible to solve signal processing tasks on high throughput data coming from multiple sources. Random Vector Functional-Link is a neural network model usually adopted in such contexts, although quantum implementations have not been considered so far. This paper proposes a quantum version of this neural model, by introducing a specific learning algorithm to find the coefficients of the adopted quantum gates and focusing on the finite precision arithmetic imposed by the qubit strings that are used to represent the model parameters.

This paper introduces a sparse learning strategy that is suited for any fuzzy inference model, in particular to the Adaptive Neuro-Fuzzy Inference System, in order to optimize the generalization capability of the resulting model. This depends on two main issues: the estimate of numerical parameters of each fuzzy rule and the whole number of rules to be used. In this work, the former problem is solved by considering a random weight fuzzy neural network, where the fuzzy rule parameters of antecedents (i.e., membership functions) are randomly generated and the ones of rule consequents are estimated using a Regularized Least Squares algorithm. The second problem is solved by pruning the coefficients of fuzzy rules following a procedure based on sparse Bayesian learning theory. Experimental results on well-known datasets prove the effectiveness of the proposed approach.

Wavelet artificial neural networks and remote sensing techniques can be used to estimate water quality variables such as Chlorophyll-a, turbidity and suspended solids. This paper describes empirical algorithms for the estimation of these variables incorporating information from the Operational Land Imager Sensor on board the Landsat-8 satellite. Neural networks are seasonally trained using data from the Cefni reservoir (Anglesey, U.K.), covering a variety of physical trophic status. Chlorophyll-a levels and the suspended solids are estimated from the reflectance in band-2 and band-4, while the turbidity values from reflectance in band-4. Experimental results show the potential of Landsat-8 as a substitute of Landsat-7 in water bodies quality monitoring. Moreover, predicted values obtained by using wavelet artificial neural networks fit well measured data and hence, such models provide accurate results therefore improving the efficiency in monitoring water quality parameters and contributing to possible decision making processes in the environmental management.

In this paper, a new algorithm for on-line learning of Random Vector Functional-Link neural network is proposed. It is specifically tailored to hardware implementations with finite precision arithmetic in distributed computing scenarios, where the massive use of low-cost hardware resources for sensor networks or computing agents is necessary in order to deal with big data, IoT paradigms and multiple sources of information. The proposed algorithm does not require any specific DSP operation to be implemented in hardware, like matrix inversions or multiplications, for real-time learning. However, experimental results prove that the algorithm outperforms commonly adopted recursive least-squares algorithms optimized for hardware implementation, while the loss of performance with respect to batch training is reduced, as the proposed approach is proved to be very efficient when a reduced number of bits is adopted for finite precision hardware implementation.

In this paper, we present a study on the use of fuzzy neural networks and their application to the prediction of times series generated by complex processes of the real-world. The new learning strategy is suited to any fuzzy inference model, especially in the case of higher-order Sugeno-type fuzzy rules. The data considered herein are real-world cases concerning chaotic benchmarks as well as environmental time series. The comparison with respect to well-known neural and fuzzy neural models will prove that our approach is able to follow the behavior of the underlying, unknown process with a good prediction of the observed time series.

In this paper, we propose a scalable, decentralized learning algorithm for Random Weights Fuzzy Neural Networks, when training data is distributed through a network of interconnected computing agents. In this scenario, the aim is for all the agents to converge to a single model, with the requirement that only local communications between the agents are permitted. In this work we assume that all the agents know the parameters of the antecedents, while the parameters of the consequents are estimated by using the Alternating Direction Method of Multipliers strategy. Experimental results show that the performance of the proposed algorithm is comparable to that of a centralized model, where all the data is collected by a single agent before the training process. To this date, this is the first publication that addressed the problem of training a fuzzy neural network over a fully decentralized infrastructure.