Previous research has demonstrated the potential of using pre-trained language models for decoding open vocabulary Electroencephalography (EEG) signals captured through a non-invasive Brain-Computer Interface (BCI). However, the impact of embedding EEG signals in the context of language models and the effect of subjectivity, remain unexplored, leading to uncertainty about the best approach to enhance decoding performance. Additionally, current evaluation metrics used to assess decoding effectiveness are predominantly syntactic and do not provide insights into the comprehensibility of the decoded output for human understanding. We present an end-to-end architecture for non-invasive brain recordings that brings modern representational learning approaches to neuroscience. Our proposal introduces the following innovations: 1) an end-to-end deep learning architecture for open vocabulary EEG decoding, incorporating a subject-dependent representation learning module for raw EEG encoding, a BART language model, and a GPT-4 sentence refinement module; 2) a more comprehensive sentence-level evaluation metric based on the BERTScore; 3) an ablation study that analyses the contributions of each module within our proposal, providing valuable insights for future research.We evaluate our approach on two publicly available datasets, ZuCo v1.0 and v2.0, comprising EEG recordings of 30 subjects engaged in natural reading tasks. Our model achieves a BLEU-1 score of 42.75%, a ROUGE-1-F of 33.28%, and a BERTScore-F of 53.86%, achieving an increment over the previous state-of-the-art by 1.40%, 2.59%, and 3.20%, respectively.

Type 1 Diabetes (T1D) affects millions worldwide, requiring continuous monitoring to prevent severe hypo- and hyperglycemic events. While continuous glucose monitoring has improved blood glucose management, deploying predictive models on wearable devices remains challenging due to computational and memory constraints. To address this, we propose a novel Lightweight Sequential Transformer model designed for blood glucose prediction in T1D. By integrating the strengths of Transformers' attention mechanisms and the sequential processing of recurrent neural networks, our architecture captures long-term dependencies while maintaining computational efficiency. The model is optimized for deployment on resource-constrained edge devices and incorporates a balanced loss function to handle the inherent data imbalance in hypo- and hyperglycemic events. Experiments on two benchmark datasets, OhioT1DM and DiaTrend, demonstrate that the proposed model outperforms state-of-the-art methods in predicting glucose levels and detecting adverse events. This work fills the gap between high-performance modeling and practical deployment, providing a reliable and efficient T1D management solution.

Telemonitoring and wearable devices are transforming healthcare by enabling continuous patient monitoring and personalized interventions. However, their integration remains fragmented due to technical, ethical, and regulatory barriers. This systematic review, following PRISMA 2020 guidelines, synthesizes insights from 165 studies (2020-2024) to assess key advancements and persistent challenges. AI-driven diagnostics show potential but require clinical validation, while energy-efficient communication protocols lack standardization. Data security improvements, such as hybrid encryption, enhance protection but introduce computational overhead. Regulatory misalignment and clinician skepticism hinder interoperability and large-scale adoption. Despite usability advancements, disparities in patient accessibility and clinician engagement highlight the need for human-centered design. To convert promises into practice, we outline a blueprint that pairs harmonized regulatory pathways with explainable, edge-deployed AI, leverages low-latency data exchange, and promotes cross-sector interoperability convergence. Realizing this agenda through longitudinal, multicenter trials and inclusive, community-anchored implementation will unlock equitable and scalable telemonitoring ecosystems.

The classification of distracted drivers is pivotal for ensuring safe driving. Previous studies demonstrated the effectiveness of neural networks in automatically predicting driver distraction, fatigue, and potential hazards. However, recent research has uncovered a significant loss of accuracy in these models when applied to samples acquired under conditions that differ from the training data. In this paper, we introduce a robust model designed to withstand changes in camera position within the vehicle. Our Driver Behavior Monitoring Network (DBMNet) relies on a lightweight backbone and integrates a disentanglement module to discard camera view information from features, coupled with contrastive learning to enhance the encoding of various driver actions. Experiments conducted using a leave-one-camera-out protocol on the daytime and nighttime subsets of the 100-Driver dataset validate the effectiveness of our approach. Cross-dataset and cross-camera experiments conducted on three benchmark datasets, namely AUCDD-V1, EZZ2021 and SFD, demonstrate the superior generalization capabilities of the proposed method. Overall DBMNet achieves an improvement of 7% in Top-1 accuracy compared to existing efficient approaches. Moreover, a quantized version of the DBMNet and all considered methods has been deployed on a Coral Dev Board board. In this deployment scenario, DBMNet outperforms alternatives, achieving the lowest average error while maintaining a compact model size, low memory footprint, fast inference time, and minimal power consumption.

Industry 5.0 envisions manufacturing systems that are human-centric, sustainable, and resilient. In this context, the Internet of Everything (IoE) enables integration of devices, people, and processes into a unified digital ecosystem. This paper presents a modular, semantically enriched framework that supports this transition by managing heterogeneous data sources—such as IoT sensors, wearable devices, and smart objects—through a layered architecture. The platform enables real-time data stream processing, semantic interoperability, and secure, context-aware access. Anomaly detection is enabled through a privacy-preserving mechanism based on behavioral fingerprinting and federated learning. The platform supports immersive human-machine interaction via gesture recognition, empowering workers to control and interact with industrial systems. Use cases demonstrate the system’s ability to support gesture-based control and intelligent monitoring, highlighting its potential to enhance adaptability, security, and worker empowerment in Industry 5.0 environments.

The use of wearable devices with inertial sensors for gesture recognition is becoming increasingly common in extended reality and remote control applications. Fast and accurate gesture recognition is critical for real-time services such as industrial control and thus drives the shift of computation toward edge computing. This work aims to optimize deep learning models for direct integration into real-world edge devices. In this context, several models based on convolutional neural networks with different configurations are developed and tested through standard performance benchmarks (accuracy and Macro F1 Score). Moreover, in order to decrease models’ size with minimal accuracy loss, they were quantized using full-integer quantization. Non-quantized and quantized models have been tested on different boards through the STM32 Edge AI Developer Cloud, as an additional benchmark to verify that the proposed models meet speed and accuracy requirements also on low-cost, low-power edge nodes. The results show that the best lightweight configuration with post-training quantization achieves an inference time of 17.31 ms on the STM32L4R9I-DISCO board, while maintaining an accuracy of 95.95 % ± 0.31 %, competitive against the state of the art and making it suitable for integration into industrial control frameworks.

Abstract State-of-the-art (SoTA) image captioning models are often trained on the MicroSoft Common Objects in Context (MS-COCO) dataset, which contains human-annotated captions with an average length of approximately ten tokens. Although effective for general scene understanding, these short captions often fail to capture complex scenes and convey detailed information. Moreover, captioning models tend to exhibit bias toward the “average” caption, which captures only the more general aspects, thus overlooking finer details. In this paper, we present a novel approach to generate richer and more informative image captions by combining the captions generated from different SoTA captioning models. Our proposed method requires no additional model training: given an image, it leverages pretrained models from the literature to generate the initial captions, and then ranks them using a newly introduced image-text-based metric, which we name BLIPScore. Subsequently, the top two captions are fused using a Large Language Model to produce the final, more detailed description. Experimental results on the MS-COCO and Flickr30k test sets demonstrate the effectiveness of our approach in terms of caption-image alignment and hallucination reduction according to the ALOHa, CAPTURE, and Polos metrics. A subjective study lends additional support to these results, suggesting that the captions produced by our model are generally perceived as more consistent with human judgment. By combining the strengths of diverse SoTA models, our method enhances the quality and appeal of image captions, bridging the gap between automated systems and the rich and informative nature of human-generated descriptions. This advance enables the generation of more suitable captions for the training of both vision-language and captioning models.

Performance in Speech Emotion Recognition (SER) on a single language has increased greatly in the last few years thanks to the use of deep learning techniques. However, cross-lingual SER remains a challenge in real-world applications due to two main factors: the first is the big gap among the source and the target domain distributions; the second factor is the major availability of unlabeled utterances in contrast to the labeled ones for the new language. Taking into account previous aspects, we propose a Semi-Supervised Learning (SSL) method for cross-lingual emotion recognition when only few labeled examples in the target domain (i.e. the new language) are available. Our method is based on a Transformer and it adapts to the new domain by exploiting a pseudo-labeling strategy on the unlabeled utterances. In particular, the use of a hard and soft pseudo-labels approach is investigated. We thoroughly evaluate the performance of the proposed method in a speaker-independent setup on both the source and the new language and show its robustness across five languages belonging to different linguistic strains. The experimental findings indicate that the unweighted accuracy is increased by an average of 40% compared to state-of-the-art methods.

Multi-modal remote sensing (RS) involves the fusion of data from multiple sensors, such as RGB, Multispectral, Hyperspectral, Light Detection and Ranging, Synthetic Aperture Radar, etc., each capturing unique information across different regions of the electromagnetic spectrum. The fusion of different modalities can provide complementary information, allowing for a comprehensive understanding of the Earth’s surface. Multi-modal RS image segmentation leverages various RS modalities to achieve pixel-level semantics classification. While deep learning has demonstrated promise in this domain, the limited availability of labeled multi-modal data poses a constraint on leveraging data-intensive techniques like deep learning to their full potential. To address this gap, we present Ticino, a novel multi-modal remote sensing dataset tailored for semantic segmentation. Ticino includes five modalities, including RGB, Digital Terrain Model, Panchromatic, and Hyperspectral images within the visual-near and short-wave infrared spectrum. Specifically annotated for Land Cover and Soil Agricultural Use, the dataset serves as a valuable resource for researchers in the field. Additionally, we conduct a comparative analysis, comparing single-modality with multi-modality deep learning techniques and evaluating the effectiveness of early fusion versus middle fusion approaches. This work aims to facilitate future research efforts in the domain by providing a robust benchmark dataset and insights into the effectiveness of various segmentation approaches.

We present a unified deep learning framework for the recognition of user identity and the recognition of imagined actions, based on electroencephalography (EEG) signals, for application as a brain–computer interface. Our solution exploits a novel shifted subsampling preprocessing step as a form of data augmentation, and a matrix representation to encode the inherent local spatial relationships of multielectrode EEG signals. The resulting image-like data are then fed to a convolutional neural network to process the local spatial dependencies, and eventually analyzed through a bidirectional long-short term memory module to focus on temporal relationships. Our solution is compared against several methods in the state of the art, showing comparable or superior performance on different tasks. Specifically, we achieve accuracy levels above 90% both for action and user classification tasks. In terms of user identification, we reach 0.39% equal error rate in the case of known users and gestures, and 6.16% in the more challenging case of unknown users and gestures. Preliminary experiments are also conducted in order to direct future works toward everyday applications relying on a reduced set of EEG electrodes.

The use of wearable devices equipped with inertial sensors has become increasingly pervasive. It has been widely demonstrated in the literature that inertial signals acquired by these sensors can be used by machine learning algorithms to predict actions performed and/or to recognize the identities of the person wearing the sensors. In this article, we present a hardware/software system for arm gesture recognition, identity recognition, and verification of a person based on inertial sensors. The hardware part is a custom wristband that consists of a computing unit, a wireless communication unit, and an inertial sensor. The software part is an algorithm based on recurrent neural networks that is able to process the signals coming from the sensor and to return a prediction. To validate the system, a dataset consisting of 25 symbols drawn with the arm is collected. These symbols are performed by 33 subjects. We conduct two evaluations: 1) performance evaluation for arm gesture recognition, user recognition and verification; and 2) usability assessment of the system. The performance of the three recognition tasks indicate that this system can be reliably applied in real environments with an accuracy above 96% for gesture recognition, an accuracy of about 85% for user identification, and an equal error rate of about 13% for user verification. The outcome of the usability test proves a great satisfaction from the users in terms of high simplicity in the use of the wristband and goodness of the machine learning predictions.

The aesthetic quality of an image is defined as the measure or appreciation of the beauty of an image. Aesthetics is inherently a subjective property but there are certain factors that influence it such as, the semantic content of the image, the attributes describing the artistic aspect, the photographic setup used for the shot, etc. In this paper we propose a method for the automatic prediction of the aesthetics of an image that is based on the analysis of the semantic content, the artistic style and the composition of the image. The proposed network includes: a pre-trained network for semantic features extraction (the Backbone); a Multi Layer Perceptron (MLP) network that relies on the Backbone features for the prediction of image attributes (the AttributeNet); a self-adaptive Hypernetwork that exploits the attributes prior encoded into the embedding generated by the AttributeNet to predict the parameters of the target network dedicated to aesthetic estimation (the AestheticNet). Given an image, the proposed multi-network is able to predict: style and composition attributes, and aesthetic score distribution. Results on three benchmark datasets demonstrate the effectiveness of the proposed method, while the ablation study gives a better understanding of the proposed network.

Current sensor-based human activity recognition techniques that rely on a user-independent model struggle to generalize to new users and on to changes that a person may make over time to his or her way of carrying out activities. Incremental learning is a technique that allows to obtain personalized models which may improve the performance on the classifiers thanks to a continuous learning based on user data. Finally, deep learning techniques have been proven to be more effective with respect to traditional ones in the generation of user-independent models. The aim of our work is therefore to put together deep learning techniques with incremental learning in order to obtain personalized models that perform better with respect to user-independent model and personalized model obtained using traditional machine learning techniques. The experimentation was done by comparing the results obtained by a technique in the state of the art with those obtained by two neural networks (ResNet and a simplified CNN) on three datasets. The experimentation showed that neural networks adapt faster to a new user than the baseline.

In the last years supervised machine learning techniques are largely employed for automatic Human Activity Recognition (HAR) using inertial sensors, such as accelerometer and gyroscope. HAR has many applications in several domains such as, for example, healthcare, sport, and entertainment. Machine learning scientists made available to the community a plenty of labeled databases for benchmarking that, unfortunately, are not consistent, both syntactically (e.g., different sampling frequency) and semantically (e.g., labels with different meanings). Commonly, due to this inconsistency, scientists evaluate their progress on individual databases separately, which corresponds to training and testing using the same database. Coherent merging of existing databases would enable: 1) evaluation of generalization capabilities of methods across databases; 2) use of deep learning techniques that, unlike traditional ones, require much more labeled data for the training process. Moreover, the growth in the daily use of wearable devices will produce a big amount of inertial data which, if not correctly labeled, cannot be efficiently exploited for the study of automatic HAR. In this paper we propose a semi-automatic procedure to coherently merge existing databases based on signal and word similarity. Preliminary experiments demonstrates the effectiveness of the proposed procedure.

In this work we present a method to improve the pruning step of the current state-of-the-art methodology to compress neural networks. The novelty of the proposed pruning technique is in its differentiability, which allows pruning to be performed during the backpropagation phase of the network training. This enables an end-to-end learning and strongly reduces the training time. The technique is based on a family of differentiable pruning functions and a new regularizer specifically designed to enforce pruning. The experimental results show that the joint optimization of both the thresholds and the network weights permits to reach a higher compression rate, reducing the number of weights of the pruned network by a further 14% to 33 % compared to the current state-of-the-art. Furthermore, we believe that this is the first study where the generalization capabilities in transfer learning tasks of the features extracted by a pruned network are analyzed. To achieve this goal, we show that the representations learned using the proposed pruning methodology maintain the same effectiveness and generality of those learned by the corresponding non-compressed network on a set of different recognition tasks.

We propose a new dataset for the evaluation of food recognition algorithms that can be used in dietary monitoring applications. Each image depicts a real canteen tray with dishes and foods arranged in different ways. Each tray contains multiple instances of food classes. The dataset contains 1027 canteen trays for a total of 3616 food instances belonging to 73 food classes. The food on the tray images has been manually segmented using carefully drawn polygonal boundaries. We have benchmarked the dataset by designing an automatic tray analysis pipeline that takes a tray image as input, finds the regions of interest, and predicts for each region the corresponding food class. We have experimented with three different classification strategies using also several visual descriptors. We achieve about 79% of food and tray recognition accuracy using convolutional-neural-networks-based features. The dataset, as well as the benchmark framework, are available to the research community.

In this article, the use of a new term extraction method for query expansion (QE) in text retrieval is investigated. The new method expands the initial query with a structured representation made of weighted word pairs (WWP) extracted from a set of training documents (relevance feedback). Standard text retrieval systems can handle a WWP structure through custom Boolean weighted models. We experimented with both the explicit and pseudorelevance feedback schemas and compared the proposed term extraction method with others in the literature, such as KLD and RM3. Evaluations have been conducted on a number of test collections (Text REtrivel Conference [TREC]‐6, ‐7, ‐8, ‐9, and ‐10). Results demonstrated that the QE method based on this new structure outperforms the baseline.

We propose a strategy for land use classification, which exploits multiple kernel learning (MKL) to automatically determine a suitable combination of a set of features without requiring any heuristic knowledge about the classification task. We present a novel procedure that allows MKL to achieve good performance in the case of small training sets. Experimental results on publicly available data sets demonstrate the feasibility of the proposed approach.

In this paper, we shall consider the problem of deploying attention to the subsets of the video streams for collating the most relevant data and information of interest related to a given task. We formalize this monitoring problem as a foraging problem. We propose a probabilistic framework to model observer's attentive behavior as the behavior of a forager. The forager, moment to moment, focuses its attention on the most informative stream/camera, detects interesting objects or activities, or switches to a more profitable stream. The approach proposed here is suitable to be exploited for multistream video summarization. Meanwhile, it can serve as a preliminary step for more sophisticated video surveillance, e.g., activity and behavior analysis. Experimental results achieved on the UCR Videoweb Activities Data Set, a publicly available data set, are presented to illustrate the utility of the proposed technique.

It is well known that supervised text classification methods need to learn from many labeled examples to achieve a high accuracy. However, in a real context, sufficient labeled examples are not always available. For this reason, there has been recent interest in methods that are capable of obtaining a high accuracy even if the size of the training set is not big. The main purpose of text mining techniques is to identify common patterns through the observation of vectors of features and then to use such patterns to make predictions. Most existing methods usually make use of a vector of features made up of weighted words that unfortunately are insufficiently discriminative when the number of features is much higher than the number of labeled examples. In this paper we demonstrate that, to obtain a greater accuracy in the analysis and revelation of common patterns, we could employ more complex features than simple weighted words. The proposed vector of features considers a hierarchical structure, named a mixed Graph of Terms, composed of a directed and an undirected sub-graph of words, that can be automatically constructed from a set of documents through the probabilistic Topic Model. The method has been tested on the top 10 classes of the ModApte split from the Reuters-21578 dataset, learned on several subsets of the original training set and showing a better performance than a method using a list of weighted words as a vector of features and linear support vector machines.

It is well known that one way to improve the accuracy of a text retrieval system is to expand the original query with additional knowledge coded through topic-related terms. In the case of an interactive environment, the expansion, which is usually represented as a list of words, is extracted from documents whose relevance is known thanks to the feedback of the user. In this paper we argue that the accuracy of a text retrieval system can be improved if we employ a query expansion method based on a mixed Graph of Terms representation instead of a method based on a simple list of words. The graph, that is composed of a directed and an undirected subgraph, can be automatically extracted from a small set of only relevant documents (namely the user feedback) using a method for term extraction based on the probabilistic Topic Model. The evaluation of the proposed method has been carried out by performing a comparison with two less complex structures: one represented as a set of pairs of words and another that is a simple list of words.

Text classification methods have been evaluated on supervised classification tasks of large datasets showing high accuracy. Nevertheless, due to the fact that these classifiers, to obtain a good performance on a test set, need to learn from many examples, some difficulties may be found when they are employed in real contexts. In fact, most users of a practical system do not want to carry out labeling tasks for a long time only to obtain a better level of accuracy. They obviously prefer algorithms that have high accuracy, but do not require a large amount of manual labeling tasks. In this paper we propose a new supervised method for single-label text classification, based on a mixed Graph of Terms, that is capable of achieving a good performance, in term of accuracy, when the size of the training set is 1% of the original. The mixed Graph of Terms can be automatically extracted from a set of documents following a kind of term clustering technique weighted by the probabilistic topic model. The method has been tested on the top 10 classes of the ModApte split from the Reuters-21578 dataset and learned on 1% of the original training set. Results have confirmed the discriminative property of the graph and have confirmed that the proposed method is comparable with existing methods learned on the whole training set.

Textual materials are source of extremely valuable information, for which there must be a reflection on the techniques of analysis to be used to avoid subjective interpretations especially in the content. The Textual Analysis (TA), which makes use of statistical techniques, ensures the systematic exploration of the structure of the text (size, occurrence, etc.) and simultaneously the possibility to return at any time to the original text for the appropriate interpretations. In this work we test a new technique based on a probabilistic model of language known in the literature as “topic model” for analyzing corpora of documents about electromagnetic pollution. The proposed method is able to reveal how the meaning of a document is distributed all along its spectrum (word-frequency) indicating that the real meaning of a document can be inferred following a multilevel analysis. Such analysis is carried out exploiting a new concept of ontology already used in literature and deeply explained here.

Mass customization is one of the most interesting and promising approach in the e-business field. In today’s competitive global market the understanding of customer needs and desires is becoming the essential preliminary remark for a successful design and implementation of products. Product development based on customer preferences with applications of innovative technologies is an essential key in order to obtain a larger market share and faster sales growth. In this scenario a tool as the product configurator is becoming a real answer to one of most important question: how to organize product design to satisfy individual customer need without trading off cost-efficiency of mass production? This paper discusses a novel approach for the design of a smart product configurator. At this moment, in fact, the configurator is just a product viewer for the customer and does not implement any reasoning logics or user adaptive approach. So an ontology based approach is presented. In this methodology three ontologies are introduced: the customer needs ontology, the product functionalities ontology and the product configuration ontology. These ontologies represent the requirement and configuration knowledge that needs for a real customization of the product. The customer has to express his product demands by the use of natural language and by the mapping among the introduced ontologies and the use of a Bayesian Network approach the automatic conversion between customer needs and product configuration is achieved.

In this paper we consider the problem of sensorimotor coordination in a Bayesian framework. To this end we introduce a novel kind of Dynamic Bayesian Network serving as the core tool to integrate active vision and task-constrained motor behaviors. The proposed system is put into work by addressing the challenging task of realistic drawing performed by a robotic agent, namely a 7-DOF anthropomorphic manipulator. Simulation results are compared to those obtained by eye-tracked human subjects involved in drawing experiments.

In this paper, we discuss how image segmentation can be handled by using Bayesian learning and inference. In particular variational techniques relying on free energy minimization will be introduced. It will be shown how to embed a spatial diffusion process on segmentation labels within the Variational Bayes learning procedure so as to enforce spatial constraints among labels.

In this paper we address the challenging problem of sensorimotor integration, with reference to eye-hand coordination of an artificial agent engaged in a natural drawing task. Under the assumption that eye-hand coupling influences observed movements, a motor continuity hypothesis is exploited to account for how gaze shifts are constrained by hand movements. A Bayesian model of such coupling is presented in the form of a novel Dynamic Bayesian Network, namely an Input–Output Coupled Hidden Markov Model. Simulation results are compared to those obtained by eye-tracked human subjects involved in drawing experiments.

We present a Bayesian model that allows to automatically generate fixations/foveations and that can be suitably exploited for compression purposes. The twofold aim of this work is to investigate how the exploitation of high-level perceptual cues provided by human faces occurring in the video can enhance the compression process without reducing the perceived quality of the video and to validate such assumption with an extensive and principled experimental protocol.

In this paper we address the issue of how form and motion can be integrated in order to provide suitable information to attentively track multiple moving objects. Such integration is designed in a Bayesian framework, and a Belief Propagation technique is exploited to perform coherent form/motion labeling of regions of the observed scene. Experiments on both synthetic and real data are presented and discussed.

Multiresolution diffused expectation maximisation performs segmentation on vector (e.g. color) images within a multiscale framework; segmentation is carried out via the expectation maximisation algorithm, coupled with anisotropic diffusion on classes, in order to account for the spatial dependencies among pixels.