Embedding neural networks in ultra-resource-constrained, low-cost, off-the-shelf microcontrollers promotes the widespread adoption of the Internet of Intelligent Things. A promising approach involves mapping network parameters from real to either binary or ternary values to reduce computational workload. Hence, we introduce Drupelet, a novel software framework for the development of mixed-precision (binary and ternary) neural networks for ultra-resource-constrained embedded devices. The modularity of Drupelet provides ease of maintenance, debugging and reuse. Drupelet was used to define the edges of memory and inference time of typical layers on two different microcontrollers. The results demonstrate that Drupelet enables mass deployment of IoIT systems, thanks to the low cost of ultra resource-constrained microcontrollers.
This paper introduces a computationally efficient approximation scheme for solving the Monge–Kantorovich (MK) optimal mass transport problem. Exact solutions for the MK problem are typically difficult or computationally expensive to obtain, particularly in high-dimensional or large-scale scenarios. To address this challenge, we propose an innovative method integrating wavelet theory and multiresolution analysis. Our approach exploits wavelet-based techniques to iteratively approximate the support of the optimal measure, thereby reducing the number of variables in linear programs and consequently decreasing the dimensionality and computational complexity of each subsequent optimization step. We present numerical experiments demonstrating that our wavelet-enhanced scheme achieves high accuracy with substantially fewer computational resources compared with traditional linear programming approaches. The method has potential applications across various domains, including image processing, economics, resource allocation, and machine learning, where efficient solutions to large-scale optimal transport problems are essential.
This paper presents SmartMathAI, an educational tool designed to support students in learning mathematics. The educational tool is based on Problem-Based Learning (PBL) with a personalized virtual tutor developed with a large language model (LLM). The virtual tutor guides the student through the problem-solving process, answers questions, and evaluates their reasoning. The use of the tool aids in mathematics learning by offering a personalized alternative that adapts to the student’s needs. SmartMathAI was evaluated with middle school students, obtaining favorable results both for learning and for student motivation when using the technology, allowing students to learn more autonomously and at their own pace.
This study presents a web module designed to predict breast cancer risk by integrating clinical and psychological variables. Recognizing that conventional models, such as Gail and Tyrer-Cuzick, primarily use biological and reproductive data, this model incorporates stress-related psychological traits, such as emotional suppression and repression, particularly relevant to the Mexican female population. Clinical and psychological data from 150 women were collected and analyzed using a novel machine learning method: Differential Evolutionary Linear Discriminant Analysis for Feature Extraction (DE-LDA_FE). The results showed that using either clinical or psychological data alone yielded moderate classification accuracy ( 62
Pouring precise liquid quantities remains a challenging task for autonomous robots. Particularly this is due to various factors as: robust container estimation, liquid’s volume estimation, arm’s or device motion uncertainty, and platform instability. This work presents a vision-based approach to estimate the tilt of cylindrical containers, a key geometric variable for inferring poured volume. The method processes frontal images using Canny edge detection and the probabilistic Hough transform to extract container and liquid boundaries. From these, a projective geometry model computes the container’s inclination without additional sensors. The proposed approach avoids the explicit calculation of fluid dynamics by providing a tool for performing a smooth pouring process. Experimental results demonstrate the feasibility of this low-cost, vision-based estimation approach.
Short-response grading is a central issue in reading comprehension evaluation. In this regard, this work first introduces an enriched dataset with both human-written and artificial intelligence–generated responses, and then compares transformer-based models, similarity metrics, and large language models for automatic validation of open-ended responses. The fine-tuned RoBERTa binary classifier achieved competitive performance, but DeepSeek-V3 outperformed all models, including ChatGPT-4o. A Sentence Transformer model trained with contrastive learning showed limitations in detecting incorrect answers. We discuss the strengths and weaknesses of each approach and propose hybrid models that are better aligned with pedagogical goals.
The increasing transformation of the European energy market, driven by the rise of intermittent renewable energies, the decommissioning of controllable power plants and dependence on short-term storage, poses challenges for assessing security of supply. In order to evaluate the capacity of available generation to meet demand in uncertain conditions, market model optimizations using the Monte Carlo (MC) approach are employed. However, the high computational costs of this approach limit assessment resolution. This paper investigates metamodeling as a strategy to reduce these computational costs. Metamodelling is a process of using mathematical models on a subset of simulations to map outcomes to the input data. This reduces the total number of simulations required. The study explores three key steps: enhancing input-output correlation, identifying effective machine learning (ML) models and selecting optimal training samples. While no single model performs adequately due to data complexity, a two-model pipeline significantly improves prediction accuracy. An active learning approach is also introduced to further optimize sample selection. The results show that training on only twenty percent of the data reduces the computation time by more than 75
Industry requires high standards of quality and precision throughout all phases of the manufacturing process, from component procurement to the assembly of final products. Visual inspection is a critical step in the quality process. It is important to note that cases such as contamination or poor-quality images in customer reports may lead to customer complaints. Therefore, the need for automatic component classification through image analysis arises as a solution. Deep learning models, and specifically convolutional neural networks, have proven to be effective tools in offering accurate and rapid image classification. This project aims to develop a simple yet efficient tool that can be easily adapted to various visual inspection tasks. Pre-trained MobileNet and ResNet models were chosen and adapted to recognize both in-focus and out-of-focus images, to distinguish between clean and dirty parts from vacuum pumps, and to test the project's flexibility in detecting good and defective cables. The goal is to make the system accurate, user-friendly, and easily adaptable. One of the advantages of this work is that anyone can update the model using just a web browser, upload or take a photo of the image to be tested and receive the prediction. This proposal aims to remove technological barriers and apply artificial intelligence to a more common area for professionals and workers in the industrial area.
This paper presents the development of a functional tool for the automatic detection and analysis of argumentative structures in Spanish academic texts. The system integrates a Conditional Random Fields (CRF) model trained on the CATyPI corpus, which consists of annotated thesis excerpts labeled with BIO tags for argumentative components. The model achieved a macro F1-score of 0.5896 and an accuracy of 0.657 using a context window size of three tokens, proving effective for identifying premises and conclusions. The system also incorporates a REST API developed with FastAPI and a user interface that visualizes the tagged text and offers real-time suggestions. These suggestions are generated by the ChatGPT 3.5 Turbo model based on the CRF model’s output to support academic writing improvement. The interface highlights argumentative elements and provides feedback that helps users refine their texts. Evaluation of the tool shows that it performs well in real use cases, although it still faces challenges with ambiguous or implicit argumentative structures. The project is publicly available through a GitHub repository and an interactive demo hosted on Hugging Face Spaces. This work demonstrates the feasibility of combining machine learning models with large language models to support argumentation in educational contexts, particularly for Spanish-speaking users.
Hate speech is a form of communication that conveys hostility, rejection, and contempt toward a person or group, based on features such as ethnic origin, religion, or gender identity. Its purpose is to foster an environment of violence and discrimination. In this paper, we present the findings of a pilot test designed to validate the dataset, the proposed methodology, and evaluate the performance of traditional algorithms in detecting ethnic hate speech. The analysis focuses on discriminatory content directed toward indigenous communities in Mexico, classified into three categories: “Hate,” “Non-Hate,” and “Unrelated Hate.” The dataset seeks to fill a gap in hate speech detection studies in Mexican Spanish, which have primarily focused on manifestations of homophobia and misogyny. We implemented traditional machine learning algorithms such as Naive Bayes, Logistic Regression, Multilayer Perceptron, and Support Vector Machine. Additionally, three different vectorizations were considered: TF-IDF, BERT, and ASCII, to identify the best way to extract features. Based on our results, we found that using the Support Vector Machine and BERT vectorization improved the classification of ethnic hate speech.
This research presents a model for detecting and counting students in educational environments. To achieve this, computer vision and computational neural networks (CNN) were used to analyze the “Classroom Monitoring Dataset.” Data cleaning and image normalization were performed with a 640 × 640 resolution. An annotation process was performed using Roboflow. Different pretrained CNN models were tested using YOLOv8 to verify the accuracy of student identification and counting. The database was divided into 80
Self-efficacy is defined as the belief of individuals in their own capabilities. Academic self-efficacy has been shown to play a crucial role in student success, it can be fostered through educational practices that enable learners to acquire skills while receiving appropriate feedback. A self-efficacy model for an intelligent learning environment is proposed to strengthen the self-efficacy of students and thus improve learning outcomes. The model predicts the self-efficacy of learners and provides pedagogical actions designed to promote it based on pedagogical models. The student’s self-efficacy model is grounded in their interaction with the learning environment, while the tutor’s self-efficacy model relies on presenting vicarious experiences, mastery experiences, and positive feedback. We hope to achieve empirical evidence on the impact of self-efficacy in adaptive learning systems, and insights about fostering beliefs of students in their own abilities in order to improve motivation, persistence, and academic achievement.
Academic writing remains a significant challenge for many university students, especially in engineering programs, where technical mastery often takes priority over communicative skills. Consequently, students frequently struggle to organize ideas, construct coherent arguments, and employ formal language appropriately. This study evaluates engineering students’ writing progress across two key dimensions: lexical richness and argumentative structure. Analyzing 69 documents from one semester, it applies Natural Language Processing (NLP) techniques to measure lexical variety, density, and sophistication, and uses a Conditional Random Fields (CRF)-based sequence-labeling model to identify premises and conclusions. Plagiarism checks and AI-generated text detection ensured authenticity. Results show notable improvements in vocabulary and argumentative clarity in some groups, with variations linked to pedagogical methods and the growing influence of generative AI. Overall, the study highlights the potential of NLP tools to provide a comprehensive assessment of academic writing in engineering education.
This paper presents a genetic algorithm (GA) for simultaneously learning the structure and parameters of Bayesian Networks (BNs) from data. The proposed method encodes both components in a single individual, using the Minimum Description Length (MDL) principle as fitness function. The algorithm is evaluated in terms of classification accuracy and complexity across different datasets. Additionally, it is assessed based on its ability to approximate known gold-standard networks. The results suggest that this approach can obtain less complex networks while preserving an acceptable level of classification accuracy and achieving a closer structural approximation to the gold standard network.
Selecting appropriate ansatz topologies for variational quantum algorithms is critical for NISQ-era success. We present a quantitative comparison between two design paradigms: the heuristic Quantum Approximate Optimization Algorithm (QAOA) and an automated Quantum Neural Architecture Search (QNAS) approach for the max-cut problem. Our analysis is based on three indicators: approximation ratio, complexity, and optimization time. The results show that QAOA achieves a 25 × complexity reduction suggests potential advantages in resource-constrained deployments—a hypothesis requiring validation under realistic noise models.
Chilacayote (Cucurbita ficifolia Bouché) has been shown to be a rich source of nutrients and bioactive compounds, suggesting its potential as a fortifier for staple foods such as corn tortillas. While fortification has been demonstrated to alter sensory properties, affecting consumer acceptance, comprehensive sensory evaluations of tortillas fortified with chilacayote powder remain scarce. In this study, sensory analyses were conducted on 150 regular tortilla consumers using CATA questionnaires to evaluate the color, smell, texture, taste, mouthfeel, and aftertaste of tortillas made from nixtamalized dough and commercial flour, both with and without chilacayote powder. The application of three feature extraction methods, Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and a combination of both (PCA+LDA), revealed that the LDA method enables the distinction of sensory differences between fortified and non-fortified tortillas. The results of this study offer valuable insights into the sensory impact of chilacayote fortification, which will inform future development of nutritionally enhanced tortillas that maintain consumer appeal.
Contrast - Limited Adaptive Histogram Equalization (CLAHE) is widely used to enhance local contrast in digital images, particularly in domains such as medical imaging, remote sensing, and low-light photography. Despite its broad adoption, the internal handling of key parameters in software libraries like OpenCV remains insufficiently documented. This work identifies that OpenCV internally rescales the user-defined contrast limit based on tile size, introducing a scaling effect not explicitly detailed in public documentation. As a result, the effective contrast-limiting behavior may diverge from both user expectations and the method’s theoretical formulation. To address this, a mathematically consistent mapping is derived that ensures the contrast limit corresponds to the number of pixels per histogram bin before clipping. A wrapper function is proposed to implement this mapping, enabling predictable and reproducible behavior across different tile configurations. The analysis contributes to both the theoretical understanding and practical application of CLAHE in image processing pipelines.
Programming skills are central to technological innovation, and intelligent learning environments are essential for acquiring them effectively. We present Javalí Notebook, a web-based interactive literate programming environment seamlessly integrated with a learning management system, designed to enhance the teaching and learning of programming. The platform allows students to engage with courses by exploring example notebooks, completing unit tests and other assessments, and incrementally developing their own programs with real-time compiler feedback and debugging support. Additionally, Javalí Notebook can serve as a classroom code presentation tool, enhancing live demonstrations and collaborative learning. The system was evaluated using the Technology Acceptance Model (TAM), and results showed positive student perceptions and intentions toward its use.
Artificial intelligence (AI) is advancing quickly, but its widespread use raises environmental concerns due to the high energy and water consumption of AI systems, particularly during training and large-scale operations. Green AI has emerged as a solution, focusing on specialized algorithms and infrastructure to lessen these environmental effects. Although the environmental footprint of standard AI methods is understood, optimization algorithms, notably Genetic Algorithms, have received less attention despite their frequent application in AI. In this study, the carbon efficiency of different combinations of crossover-mutation operators in a real genetic algorithm (rGA) was examined. The emissions were measured across a range of settings, finding consistent patterns in emissions ranging from 2.669E-05 to 2.204E-04 kg CO _2 eq. While small in scale, these values can add up to significant costs in larger experimental or real-world scenarios. Statistical tests confirmed significant differences; combinations such as 1P_LM (one-point crossover and boundary mutation), 2P_LM (two-point and boundary mutation), and UNI_LM (uniform crossover and boundary mutation) consistently yielded the lowest emissions.
This study presents a benchmark for evaluating visual correspondence algorithms in underwater environments using both optical and sonar imaging. It analyzes the transferability of state-of-the-art feature matching methods designed initially for terrestrial data, under the specific challenges of marine sensing. Experiments on real and simulated datasets assess their accuracy, robustness, and downstream impact on visual odometry and image mosaicing. The findings highlight key limitations in generalization and provide insights toward developing more trustworthy perception systems for autonomous underwater robots (This work was supported in part by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - project number 535678995.).
