The main contribution is the design, implementation and validation of a complete AI-driven electronic nose architecture to perform the classification of whiskey and acetones. This classification is of paramount important in the distillery production line of whiskey in order to predict the quality of the final product. In this work, we investigate the application of an e-nose (based on arrays of single-walled carbon nanotubes) to the distinction of two different substances, such as whiskey and acetone (as a subproduct of the distillation process), and discrimination of three different types of the same substance, such as three types of whiskies. We investigated different strategies to classify the odor data and provided a suitable approach based on random forest with accuracy of 99% and with inference times under 1.8 seconds. In the case of clearly different substances, as subproducts of the whiskey distillation process, the procedure presented achieves a high accuracy in the classification process, with an accuracy around 96%.

Conversational Intelligent Tutoring Systems (CITS) have drawn increasing interest in education because of their capacity to tailor learning experiences, improve user engagement, and contribute to the effective transfer of knowledge. Conversational agents employ advanced natural language techniques to engage in a convincing human-like tutorial conversation. In solving math word problems, a significant challenge arises in enabling the system to understand user utterances and accurately map extracted entities to the essential problem quantities required for problem-solving, despite the inherent ambiguity of human natural language. In this study, we propose two possible approaches to enhance the performance of a particular CITS designed to teach learners to solve arithmetic–algebraic word problems. Firstly, we propose an ensemble approach to intent classification and entity extraction, which combines the predictions made by two distinct individual models that use constraints defined by human experts. This approach leverages the intertwined nature of the intents and entities to yield a comprehensive understanding of the user’s utterance, ultimately aiming to enhance semantic accuracy. Secondly, we introduce an adapted Term Frequency-Inverse Document Frequency technique to associate entities with problem quantity descriptions. The evaluation was conducted on the AWPS and MATH-HINTS datasets, containing conversational data and a collection of arithmetical and algebraic math problems, respectively. The results demonstrate that the proposed ensemble approach outperforms individual models, and the proposed method for entity–quantity matching surpasses the performance of typical text semantic embedding models.

Despite the general consensus on the advantages of collaborative learning, most of the research in Intelligent Tutoring Systems (ITSs) only considers the case of individual learners. This is mainly due to the technical challenges in adapting individual learning systems to support collaborative learning. In this paper, we present a framework aimed at adapting individual-learner tutoring systems to enable collaboration among students, leveraging the advantages of both intelligent tutoring systems and peer collaboration, without significant changes to the original Human–Computer Interaction model. The proposed framework is designed under the assumption that the adapted ITS is built as a web application, ensures horizontal scalability of the ITS, and relies exclusively on Open Source software and tools. Evaluation results of the framework demonstrate that while the system’s complexity increases, there are no perceivable rises in response times or cpu usage. • We present a framework for adapting Intelligent Tutoring Systems to collaborative. • Reduced technical challenge for implementing Collaboration. • We provide a case study of the adaptation of the system. • The modified system is not perceivably slower.

The performance of Conversational Agents (CAs) relies heavily on intent classification and entity extraction. However, the effectiveness of these tasks is often hindered by the absence of an explicit model that accounts for their interdependence. To address this limitation, we introduce a novel approach that leverages the inter-dependency between intents and entities to improve the Exact Match Accuracy (EMA) of CAs. The approach evaluates the consistency of the output generated by the Natural Language Understanding (NLU) component based on a specification that outlines all valid combinations of intents and entities. If an inconsistent output is detected, the ranking of predicted intents is leveraged to determine the most probable intent that aligns with the identified entities. The technique guarantees the preservation or improvement of EMA and can be applied as a post-processing step in combination with any existing NLU method that returns a ranking of intents. The proposed approach was evaluated through an ablation study conducted using three different NLU models (DIET, DCA-Net and Bi-model) on four distinct conversational datasets: ATIS, SNIPS, NLU-Benchmark, and AWPS. The results demonstrated that our method led to a consistent improvement in the EMA across all NLU methods and datasets. The magnitude of the improvement varies depending on the method, dataset, and training data size, ranging from below 1% when using DCA-Net or Bi-model on the SNIPS dataset with full training data, to over 9% when using Bi-model or DIET on AWPS with 10% of the available training data. • A novel approach to increase the exact match accuracy in NLU systems is presented. • The method is non-deteriorating and seamlessly integrates with existing NLU systems. • The technique leverages the inherent inter-dependency between intents and entities. • It amends incompatible outputs by using a human-provided constraint specification. • The system’s performance consistently improved in 4 commonly used NLU datasets.

Identification of animal species is a crucial aspect of biology and ecology, particularly in ornithology, where collaboration with other disciplines aims to develop effective methods for bird protection and environmental quality assessment. Leveraging artificial intelligence (AI) and Internet of Things (IoT) technologies, advancements in birdsong identification have been achieved. Machine learning and deep learning techniques, including Imagenet-based Convolutional Neural Networks (CNNs) like EfficientNet and MobileNet, have been employed for image feature comparison. Spectrograms of birdsongs have been analyzed, with Deep CNNs (DCNNs) proving effective in reducing model size for birdsong classification.

Recent Large Language Models (LLMs) demonstrate problem-solving capabilities suitable for educational use. This paper investigates using LLMs to create virtual agents that mimic student behavior and interact with learning platforms. Testing a modest-sized LLM on an Intelligent Tutoring System for word problem-solving revealed that the LLM could fully solve 92% of single-step problems, although their performance decreased to 14% when attempting more complex problems.

Recent advances in generative AI epitomized by Large Language Models (LLMs) demonstrated remarkable capabilities in generating human-like text and understanding contextual nuances. In parallel, advances in domestic applications of AI continue to break new grounds bringing to the forefront the need to address the impact of such systems especially on users' well being through socially aware interactive intelligent systems. This paper takes one step forward in this endeavour to examine the potential of LLMs based systems to generate empathy in its responses. In doing so, we identified through initial tests the key factors to designing empathy triggering responses. We then designed a set of experiments to test systematically the impact each of these factors would have on triggering empathy. We analysed the results against the degree of expressed empathy and its perceived quality.

Abstract Natural Language Understanding (NLU) components are used in Dialog Systems (DS) to perform intent detection and entity extraction. In this work, we introduce a technique that exploits the inherent relationships between intents and entities to enhance the performance of NLU systems. The proposed method involves the utilization of a carefully crafted set of rules that formally express these relationships. By utilizing these rules, we effectively address inconsistencies within the NLU output, leading to improved accuracy and reliability. We implemented the proposed method using the Rasa framework as an NLU component and used our own conversational dataset AWPS to evaluate the improvement. Then, we validated the results in other three commonly used datasets: ATIS, SNIPS, and NLU-Benchmark. The experimental results show that the proposed method has a positive impact on the semantic accuracy metric, reaching an improvement of 12.6% in AWPS when training with a small amount of data. Furthermore, the practical application of the proposed method can easily be extended to other Task-Oriented Dialog Systems (T-ODS) to boost their performance and enhance user satisfaction.

Remote photoplethysmography (rPPG) has gained attention as a non-intrusive camera-based contactless method for the remote monitoring of vital signs. rPPG is especially appealing for its suitability in telehealth settings as a way to provide healthcare professionals physiological measurements of their patients during video consultations, without the need of specialist equipment. Limited data transmission is a key concern in video conferencing and necessitates the need for video compression to mitigate the communication bottleneck. To ensure that their service is manageable at large scales, video communication providers often use a form of Constant Bitrate (CBR) mode encoding to keep bitrates constrained and consistent regardless of video content. In this paper, we study how video compression using Constant Bitrate mode encoding affects the accuracy of rPPG methods and compare this to video encoded using a Constant Rate Factor mode (CRF), more commonly used for asynchronous video. To this end, we analyse the behaviour of three typically used video codecs (AV1, H.264, and H.265) in both CRF and CBR modes on 129 videos from 3 different databases (UBFC-RPPG, LGI-PPGI-FVD and PURE). Our results show a noticeable difference between CRF and CBR mode encoded videos on rPPG signal quality that is inconsistent between encoders.

In this paper, we present and evaluate the recent incorporation of a conversational agent into an Intelligent Tutoring System (ITS), using the open-source machine learning framework Rasa. Once it has been appropriately trained, this tool is capable of identifying the intention of a given text input and extracting the relevant entities related to the message content. We describe both the generation of a realistic training set in Spanish language that enables the creation of the required Natural Language Understanding (NLU) models and the evaluation of the resulting system. For the generation of the training set, we have followed a methodology that can be easily exported to other ITS. The model evaluation shows that the conversational agent can correctly identify the majority of the user intents, reporting an f1-score above 95%, and cooperate with the ITS to produce a consistent dialogue flow that makes interaction more natural.

Accurately assessing task difficulty is a critical aspect to achieve adaptation in computer-based educational systems. In real-world scenarios, task difficulty estimation can be personalised for individuals by leveraging Item Response Theory (IRT) to analyse the collective performance of a group of students across various tasks. Additionally, recent studies have revealed the potential of inferring task difficulty through the analysis of physiological signals, such as electrocardiography (ECG). In this paper, we propose a novel hybrid approach that combines both methodologies to enhance task difficulty estimates, surpassing the individual performance of each method. The availability of non-intrusive techniques for capturing heart rate adds further value to the proposal, facilitating its potential integration into future computer-based educational systems. Experimental results on a dataset captured during two computerised English tests show that our proposed hybrid approach outperforms each individual method for the task of difficulty estimation.

Conversational Intelligent Tutoring Systems (CITS) in learning environments are capable of providing personalized instruction to students in different domains, to improve the learning process. This interaction between the Intelligent Tutoring System (ITS) and the user is carried out through dialogues in natural language. In this study, we use an open source framework called Rasa to adapt the original button-based user interface of an algebraic/arithmetic word problem-solving ITS to one based primarily on the use of natural language. We conducted an empirical study showing that once properly trained, our conversational agent was able to recognize the intent related to the content of the student’s message with an average accuracy above 0.95.

The availability of low-cost wireless physiological sensors has allowed the use of emotion recognition technologies in various applications. In this work, we describe a technique to predict emotional states in Russell’s two-dimensional emotion space (valence and arousal), using electroencephalography (EEG), electrocardiography (ECG), and electromyography (EMG) signals. For each of the two dimensions, the proposed method uses a classification scheme based on two Hidden Markov Models (HMMs), with the first one trained using positive samples, and the second one using negative samples. The class of new unseen samples is then decided based on which model returns the highest score. The proposed approach was validated on a recently published dataset that contained physiological signals recordings (EEG, ECG, EMG) acquired during a human-horse interaction experiment. The experimental results demonstrate that this approach achieves a better performance than the published baseline methods, achieving an F1-score of 0.940 for valence and 0.783 for arousal, an improvement of more than + 0.12 in both cases.

Various recent research works have focused on the use of electroencephalography (EEG) signals in the field of biometrics. However, advances in this area have somehow been limited by the absence of a common testbed that would make it possible to easily compare the performance of different proposals. In this work, we present a data set that has been specifically designed to allow researchers to attempt new biometric approaches that use EEG signals captured by using relatively inexpensive consumer-grade devices. The proposed data set has been made publicly accessible and can be downloaded from https://doi.org/10.5281/zenodo.4309471. It contains EEG recordings and responses from 21 individuals, captured under 12 different stimuli across three sessions. The selected stimuli included traditional approaches, as well as stimuli that aim to elicit concrete affective states, in order to facilitate future studies related to the influence of emotions on the EEG signals in the context of biometrics. The captured data were checked for consistency and a performance study was also carried out in order to establish a baseline for the tasks of subject verification and identification.

Biometric signals have been extensively used for user identification and authentication due to their inherent characteristics that are unique to each person. The variation exhibited between the brain signals (EEG) of different people makes such signals especially suitable for biometric user identification. However, the characteristics of these signals are also influenced by the user's current condition, including his/her affective state. In this paper, we analyze the significance of the affect-related component of brain signals within the subject identification context. Consistent results are obtained across three different public datasets, suggesting that the dominant component of the signal is subject-related, but the affective state also has a contribution that affects identification accuracy. Results show that identification accuracy increases when the system has been trained with EEG recordings that refer to similar affective states as the sample that is to be identified. This improvement holds independently of the features and classification algorithm used, and it is generally above 10 percent under a rigorous setting, when the training and validation datasets do not share data from the same recording days. This finding emphasizes the potential benefits of considering affective information in applications that require subject identification, such as user authentication.

The increasing popularity of video applications and ever-growing high-quality video transmissions (e.g., 4K resolutions), has encouraged other sectors to explore the growth of opportunities. In the case of health sector, mobile Health services are becoming increasingly relevant in real-time emergency video communication scenarios where a remote medical experts' support is paramount to a successful and early disease diagnosis. To minimize the negative effects that could affect critical services in a heavily loaded network, it is essential for 5G video providers to deploy highly scalable and priorizable in-network video optimization schemes to meet the expectations of a large quantity of video treatments. This paper presents a novel 5G Video Optimizer Virtual Network Function (vOptimizerVNF) that leverages the latest technologies in 5G and video processing to address this important challenge. Advanced traffic filtering is coupled with Scalable H.265 video coding to enable run-time bandwidth-saving video optimization without compromising Quality of Service (QoS); kernel-space video processing is introduced to achieve further performance gains; and the use of a Virtual Network Function (VNF) facilitates dynamic deployment of virtualized video optimizers to achieve scalability and flexibility in this service. The proposed approach is implemented in a realistic 5G testbed and empirical results demonstrate the superior scalability and performance achieved.

Electroencephalography (EEG) signals provide a representation of the brain's activity patterns and have been recently exploited for user identification and authentication due to their uniqueness and their robustness to interception and artificial replication. Nevertheless, such signals are commonly affected by the individual's emotional state. In this work, we examine the use of images as stimulus for acquiring EEG signals and study whether the use of images that evoke similar emotional responses leads to higher identification accuracy compared to images that evoke different emotional responses. Results show that identification accuracy increases when the system is trained with EEG recordings that refer to similar emotional states as the EEG recordings that are used for identification, demonstrating an up to 5.3% increase on identification accuracy compared to when recordings referring to different emotional states are used. Furthermore, this improvement holds independently of the features and classification algorithms employed.

The relation between affect-driven feedback and engagement on a given task has been largely investigated. This relation can be used to make personalised instructional decisions and/or modify the affect content within the feedback. However, although it is generally assumed that providing encouraging feedback to students should help them adopt a state of flow, there are instances where those messages might result counterproductive. In this paper, we present a case study with 48 secondary school students using an Intelligent Tutoring System for arithmetical word problem solving. This system, which makes some common assumptions on how to relate affective state with performance, takes into account subjective (user's affective state) and objective information (previous problem performance) to decide the upcoming difficulty levels and the type of affective feedback to be delivered. Surprisingly, results revealed that feedback was more effective when no emotional content was included, and lead to the conclusion that purely instructional and concise help messages are more important than the emotional reinforcement contained therein. This finding shows that this is still an open issue. Different settings present different constraints generating related compounding factors that affect obtained results. This research confirms that new approaches are required to determine when, how and where affect-driven feedback is needed. Affect-driven feedback, engagement and their mutual relation have been largely investigated. Student's interactions combined with their emotional state can be used to make personalised instructional decisions and/or modify the affect content within the feedback, aiming to entice engagement on the task. However, although it is generally assumed that providing encouraging feedback to the students should help them adopt a state of flow, there are instances where those encouraging messages might result counterproductive. In this paper, we analyze these issues in terms of a case study with 48 secondary school students using an Intelligent Tutoring System for arithmetical word problem solving. This system, which makes some common assumptions on how to relate affective state with performance, takes into account subjective (user's affective state) and objective (previous problem performance) information to decide the difficulty level of the next exercise and the type of affective feedback to be delivered. Surprisingly, findings revealed that feedback was more effective when no emotional content was included in the messages, and lead to the conclusion that purely instructional and concise help messages are more important than the emotional reinforcement contained therein. This finding, which coincides with related work, shows that this is still an open issue. Different settings present different constraints and there are related compounding factors that affect obtained results, such as the message's contents and their target, how to measure the effect of the message on engagement through affective variables considering other issues involved, and to what extent engagement can be manipulated solely in terms of affective feedback. The contribution here is that this research confirms that new approaches are needed to determine when, how and where affect-driven feedback is needed. In particular, based on our previous experience in developing educational recommender systems, we suggest the combination of user-centred design methodologies with data mining methods to yield a more effective feedback.

Security systems are starting to meet new technologies and new machine learning techniques, and a variety of methods to identify individuals from physiological signals have been developed. In this paper, we present ESID, a deep learning approach to identify subjects from electroencephalogram (EEG) signals captured by using a low cost device. The system consists of a Convolutional Neural Network (CNN), which is fed with the power spectral density of different EEG recordings belonging to different individuals. The network is trained for a period of one million iterations, in order to learn features related to local patterns in the spectral domain of the original signal. The performance of the system is evaluated against other traditional classification-based methods that use prior-knowledge-defined features. Results show that the system significantly outperforms other examined approaches, with 94% accuracy at discerning an individual in between a group of 23 different individuals.

Despite many research efforts focused on the development of algebraic reasoning and the resolution of story problems, several investigations have reported that relatively advanced students experience serious difficulties in symbolizing certain meaningful relations by using algebraic equations. In this paper, we describe and justify the Graphical User Interface of an Intelligent Tutoring System that allows learning and practising the procedural aspects involved in translating the information contained in a story problem into a symbolic representation. The application design has been driven by cognitive findings from several previous investigations. First, the process of translating a word problem into an algebraic form has been treated in isolation, and clearly separated from algebraic manipulation. Second, the user interface has been devised to force a systematic approach to problem solving, and also avoid the use of a non-algebraic reasoning. Third, sensor-free affective support has been added by using a machine learning approach that relies on data captured from a series of experimental sessions involving 48 subjects. The evaluation of the resulting application has revealed a positive and significant impact in learning gains.

The selection of an appropriate sequence of activities is an essential task to keep student motivation and foster engagement. Usually, decisions in this respect are made by taking into account the difficulty of the activities, in relation to the student's level of competence. In this paper, we present a dynamic model that aims to predict the average performance of a group of students at solving a given series of maths problems. The system takes into account both student- and task-related features. This model was built and validated by using the data gathered in an experimental session that involved 64 participants solving a sequence of 26 arithmetic problems. The data collected from the first 16 problems was used to build the model, and the remainder were employed to validate it. Results show a correlation with r = 0.59 between the real and predicted scores, and support the effectiveness of the model at anticipating the students' performance over a sequence of tasks.

Most previous work in emotion recognition has fixed the available classes in advance, and attempted to classify samples into one of these classes using a supervised learning approach. In this paper, we present preliminary work on combining supervised and unsupervised learning to discover potential latent classes which were not initially considered. To illustrate the potential of this hybrid approach, we have used a Self-Organizing Map (SOM) to organize a large number of Electroencephalogram (EEG) signals from subjects watching videos, according to their internal structure. Results suggest that a more useful labelling scheme could be produced by analysing the resulting topology in relation to user reported valence levels (i.e., pleasantness) for each signal, refining the original set of target classes.

One basic adaptation function of an Intelligent Tutoring System (ITS) consists of selecting the most appropriate next task to be offered to the learner. This decision can be based on estimates, such as the expected performance of the student, or the probability that the student successfully solves each particular task. However, the computation of these values is intrinsically difficult, as they may depend on other complex latent variables that also need to be estimated from observable quantities, e.g. the current student's ability. In this work, we have used system dynamics to model learning and predict the student's performance in a given exercise, in an existing ITS that was developed to teach students solve arithmetic-algebraic word problems. The high correlation between the predicted and real scores outlines the potential of this type of modeling as a prediction tool to support the decision about the next task that should be offered to the learner.

There exist a large number of distance functions that allow one to measure similarity between feature vectors and thus can be used for ranking purposes. When multiple representations of the same object are available, distances in each representation space may be combined to produce a single similarity score. In this paper, we present a method to build such a similarity ranking out of a family of distance functions. Unlike other approaches that aim to select the best distance function for a particular context, we use several distances and combine them in a convenient way. To this end, we adopt a classical similarity learning approach and face the problem as a standard supervised machine learning task. As in most similarity learning settings, the training data are composed of a set of pairs of objects that have been labeled as similar/dissimilar. These are first used as an input to a transformation function that computes new feature vectors for each pair by using a family of distance functions in each of the available representation spaces. Then, this information is used to learn a classifier. The approach has been tested using three different repositories. Results show that the proposed method outperforms other alternative approaches in high-dimensional spaces and highlight the benefits of using multiple distances in each representation space.

The vast majority of localization approaches using phase transform (PHAT) consider that the sources of interest are wideband low-pass sources. While this may be the usual case for common audio signals such as speech, PHAT methods are affected negatively by modulation artifacts when the sources to be localized are passband signals. In these cases, steered response power PHAT localization becomes less robust. This letter analyzes the form of generalized cross-correlation functions with PHAT when passband acoustic signals are considered, proposing approaches for increasing the localization performance through the mitigation of these negative effects.

In this paper, we present an effective algorithm to reduce the number of wraps in a 2D phase signal provided as input. The technique is based on an accurate estimate of the fundamental frequency of a 2D complex signal with the phase given by the input, and the removal of a dependent additive term from the phase map. Unlike existing methods based on the discrete Fourier transform (DFT), the frequency is computed by using noise-robust estimates that are not restricted to integer values. Then, to deal with the problem of a non-integer shift in the frequency domain, an equivalent operation is carried out on the original phase signal. This consists of the subtraction of a tilted plane whose slope is computed from the frequency, followed by a re-wrapping operation. The technique has been exhaustively tested on fringe projection profilometry (FPP) and magnetic resonance imaging (MRI) signals. In addition, the performance of several frequency estimation methods has been compared. The proposed methodology is particularly effective on FPP signals, showing a higher performance than the state-of-the-art wrap reduction approaches. In this context, it contributes to canceling the carrier effect at the same time as it eliminates any potential slope that affects the entire signal. Its effectiveness on other carrier-free phase signals, e.g., MRI, is limited to the case that inherent slopes are present in the phase data.

Quality-based 2D phase unwrapping algorithms provide one of the best tradeoffs between speed and quality of results. Their robustness depends on a quality map, which is used to build a path that visits the most reliable pixels first. Unwrapping then proceeds along this path, delaying unwrapping of noisy and inconsistent areas until the end, so that the unwrapping errors remain local. We propose a novel quality measure that is consistent, technically sound, effective, fast to compute, and immune to the presence of a carrier signal. The new measure combines the benefits of both the quality-guided and the residue-based phase unwrapping approaches. The quality map is justified from the two different theoretical points of view. Exhaustive tests on a variety of artificially generated and real 2D wrapped phase signals illustrate its potential usefulness in the field of fringe projection profilometry.

Distance metric learning aims at obtaining an appropriate metric that conveniently adapts to a particular recognition problem given a set of training pairs. The idea of maximizing a margin that separates similar and dissimilar objects has been used in different ways in several recent works. This paper considers two different learning schemes aiming at the same goal but posing the learning problem either as a batch or as an online formulation. Extensive experiments and the corresponding discussion try to put forward the advantages and drawbacks of each of the approaches considered.

The advent of the Internet has opened a scope for research in new methods and tools that may facilitate the teaching and learning processes. This has, in turn, led to the development of learning platforms to support teaching and learning activities. The market penetration of these has been such that nowadays most universities provide their academic community with some form of a learning management system (LMS).Although a lot of effort has been put into deploying these platforms, the usage statistics that they generally provide are not generally processed to optimize their use within a specific context or institution (or confronted with quality or innovation indexes to produce useful feedback information). In this paper we propose a methodology for data analysis and apply it to the particular case of the University of Valencia(Spain). Besides, we briefly describe a computer implementation that facilitates the application of this methodology to any other existing LMS which provides usage statistics.

Content based image retrieval (CBIR) systems aim to provide a means to find pictures in large repositories without using any other information except its contents usually as low-level descriptors. Since these descriptors do not exactly match the high level semantics of the image, assessing perceptual similarity between two pictures using only their feature vectors is not a trivial task. In fact, the ability of a system to induce high level semantic concepts from the feature vector of an image is one of the aspects which most influences its performance. This paper describes a CBIR algorithm which combines relevance feedback, evolutionary computation concepts and ad-hoc strategies in an attempt to fill the existing gap between the high level semantic content of the images and the information provided by the low level descriptors.

Image retrieval has become an interesting and active field due to the increasing necessity of searching and browsing very large image repositories. Images are represented using several kinds of low-level descriptors from which convenient similarity or score functions are computed. Recent work deals with different ways of combining these measures to improve the overall performance of the retrieval system. This paper builds upon previous ideas taken from different contexts to deploy a convenient combination framework that takes into account learning data directly gathered from the users that are supposed to end using the system. The proposal is empirically evaluated and compared to other ways of combining the same measures.