Scholar dropout is a phenomenon that affects many Higher Education Institutions in Mexico. Economic problems and failure are the main causes of this problem; however, through traditional machine learning techniques, many authors have approached this study. In this paper, the study of scholar dropout in Technological University in Puebla will be approached using deep learning techniques that are commonly used in the field of image classification and natural language processing. In particular, deep neural network techniques such as convolutional neural networks are considered, performing a treatment of the data set used to emulate the data structures required by these types of algorithm. Finally, an analysis of the performance of these models and their relevance in the study of school dropout is carried out.
Inductive biases are essential for efficient inductive learning. In neural networks, convolutional layers are biased towards learning location-invariant features, whereas attention layers excel at capturing dependencies in sequential data. Here, we systematically study and unify the biases of these architectures within the framework of geometric deep learning to reveal their underlying similarities. We demonstrate that local attention layers exhibit a relational inductive bias similar to that of convolutional layers. Based on this finding, we construct a local attention layer for image processing that assumes a grid structure but models the data dynamically and stochastically, in contrast to conventional convolutional layers, which use static and deterministic assumptions.
This work presents the "Data-driven Accumulated Nonlinear Imputation" (DANLI) method that combines energy time series consumption profiles with pattern-matching to reconstruct missing values. It is designed for scenarios with only boundary measurements and is based on the hypothesis that cumulative power and daily usage similarities matter more than instantaneous power consumption in many applications. Results indicate that compared to statistical imputation, interpolation, and point-wise Euclidean pattern-matching, DANLI generates physically plausible reconstructions while achieving competitive accuracy across gaps.
Computed Tomography (CT) is a widely used imaging technique, but its reliance on ionizing radiation limits its applicability in repeated or sensitive medical/industrial scenarios. Ultrasound Computed Tomography (USCT) offers a safer alternative; however, the resulting reconstructions are often of low resolution and prone to artifacts, making anomaly detection challenging. In this work, we propose a two-stage deep learning pipeline (DLP) for anomaly classification and localization in simulated USCT images. Numerical experiments consider a viscoelastic material immersed in water with properties similar to human tissue. Training and validation datasets were primarily generated from synthetic handwritten images reconstructed via backprojection tomography (BP) and complemented with a small subset ( ≈ 6% ) of numerically simulated ultrasonic tomography reconstruction (FBP-USCT). In the first stage of DLP, a ResNet-50 model addresses a six-class classification task, distinguishing between images with and without anomalies, achieving 90.16% accuracy. In the second stage, a modified Faster R-CNN with VGG-16 as backbone is used for anomaly localization, reaching 85.65 87% precision on a test set excluded from training and validation. These results is a proof-of-concept that the proposed CNN-based model can effectively perform the classification and detection tasks, thereby reducing the cost of numerical or experimental datasets of ultrasonic tomography.
This paper introduces a compact hexapod that combines a single USB camera with two front-mounted ultrasonic sensors to perform fully autonomous target localization indoors. Eighteen smart servos are coordinated by an ESP32, while a Raspberry Pi 5 runs a 10 FPS YOLOv1l detector whose centroid and confidence values feed the influence term of a Repulsion-Attraction-Orientation-Influence (RAOI) controller; the ultrasonic sweep generated by the moving front legs supplies the repulsion term, so no side sonars or depth sensors are required. Early tests showed that a tripod gait introduced oscillations on the 1.5 kg platform, so the controller was switched to a wave gait that keeps at least four legs grounded, yielding smoother motion for vision-based steering. The robot was evaluated in a 2.3 × 3.2 m. arena with a 0.6 L bottle placed at azimuths of 0^∘ , 45^∘ and 90^∘ . Across fifteen replicas (five per azimuth) detection precision was 100
The effectiveness of cross-lingual transfer in Transformer models is critical for applying NLP to low-resource languages, yet its practical boundaries remain underexplored. This work investigates this phenomenon through a case study on sensitive information classification in Portuguese, explicitly mapping the limits and surprises of cross-lingual adaptation. We systematically compare four BERT variants: a specialized Portuguese model (BERTimbau), a multilingual model (mBERT), and two monolingual models from distant languages (English and Chinese). Our findings reveal that while the specialized model’s superiority (99.49
Fear of public speaking is one of the most prevalent forms of social anxiety, significantly impairing academic and professional performance. This study presents the development of a prototype virtual reality (VR) treatment supported by artificial intelligence (AI), aimed at reducing social anxiety in academic contexts. A decision tree model was trained using the C4.5 supervised learning algorithm on data from 134 university students. Based on the model’s classifications, immersive VR environments simulating public speaking scenarios were designed and implemented using the Unity engine and C#. User experience was evaluated through the NASA Task Load Index (NASA-TLX) in a sample of 20 psychology students. The model achieved 97
This paper presents the Stochastic Harmonic Potential Field (SHPF) method for provably safe navigation of differential-drive mobile robots in dynamic environments using 2D LiDAR sensing. SHPF combines harmonic potential fields, velocity-adaptive stochastic prediction of obstacle motion, and an adaptive control law to ensure local-minima-free trajectories with real-time safety guarantees. Theoretical analysis establishes sufficient conditions for collision avoidance under maximum velocity constraints, while extensive simulations demonstrate superior performance over five state-of-the-art methods (DWA, VO, ORCA, RRT*, MPC). Evaluated in two benchmark scenarios—dynamic corridor navigation and cross-traffic avoidance—SHPF achieved a 100
Tomosynthesis is a cutting-edge medical imaging technique that utilizes multiple X-ray projections to create high-resolution, threedimensional images. Conversely, sinograms constitute two-dimensional representations derived from the scanning process, which concentrate the largest amount of information but suffer from noise and low contrast, complicating accurate interpretation. This study compares two methodologies for enhancing sinograms: a traditional histogram-based contrast adjustment method and a more innovative contrast enhancement approach utilizing a Particle Swarm Optimization (PSO) algorithm. The conventional method focuses on adjusting contrast and smoothing images to mitigate noise. At the same time, the PSO approach aims to identify optimal parameters for a transformation function that effectively enhances contrast. Both techniques were rigorously evaluated using established metrics for contrast enhancement, including Peak Signalto-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Feature Similarity Index Measure (FSIM). The findings reveal that, although the traditional method performs adequately for simpler images, the approach based on PSO significantly enhances the quality of more complex sinograms, particularly those impacted by artifacts and elevated noise levels. This research thoroughly examines current methodologies for improving sinogram quality, highlighting their potential applications in medical diagnostics and advancements in imaging technology.
Technological advances in computer vision have moved from detecting and recognizing objects in an image, to understanding the relationships between objects and generating a textual description based on the image content, using an image dataset and scene graphs. Scene graphs are a structured representation of a scene that can express the objects, attributes and relationships between objects in the scene, in order to textually describe the image content. In this research work, we present the creation of a dataset of bee images called “Bees” and the generation of scene graphs from this dataset, in order to recognize, classify and analyze the organizational behavior of bees at the entrance of the beehive, for the sake of the preservation and care of this endangered species. We created an image dataset of 100 images and their constituent parts necessary for the generation of the scene graphs. The method used is described and evaluated with a data set of 500 images. In the generation of scene graphs, the experiments performed show on average 69
The Multi-Objective Evolutionary Algorithm based on Decomposition (MOEA/D) has emerged as a robust and computationally efficient framework for addressing complex optimization challenges. In recent years, there has been a significant focus on adapting MOEA/D to effectively tackle constrained multi-objective optimization problems. This paper introduces a novel enhancement to the MOEA/D framework through the integration of a dynamic penalty function aimed at improving constraint handling. To evaluate the effectiveness of our proposed approach, we conduct extensive experiments using the widely recognized CEC'2009 benchmark problems. Our methodology is rigorously compared against state-of-the-art multi-objective optimization algorithms, allowing for a comprehensive assessment of its performance. The experimental results show that our enhanced MOEA/D approach yields solutions that are not only competitive but, in specific instances, outperform those generated by the leading algorithms in the field. Additionally, we discuss the implications of our findings for future research and practical applications, highlighting the potential of our dynamic penalty function to advance the state of the art in constrained multi-objective evolutionary optimization.
Background: Childhood overweight and obesity are increasingly prevalent, particularly in low- and middle-income countries, and predispose individuals to chronic diseases. Conventional clinical assessments often overlook the multifactorial nature of obesity, leading to generalized and less effective interventions. This study evaluated the effectiveness of an AI ensemble model to predict school-age nutritional status from preschool data in Mexico. Methods: Data were obtained from the 2018 Mexican National Health and Nutrition Survey, including 8,327 healthy children under 5 years. A total of 75 clinical, demographic, socioeconomic, perinatal and behavioral variables were collected. The prediction task was framed as a four-class problem using WHO weight-for-height z-score. Missing values were imputed with a random forest method within the cross-validation folds to prevent data leakage. Three base models (multinomial logistic regression, support vector machine, and random forest) were trained, optimized and stacked into a two-hidden-layer neural network. Performance was assessed using accuracy, precision, F1-score, and AUC, reported as macro-averages across the four classes. Model interpretability was examined using LIME. Results: Class distribution at age five was imbalanced (undernutrition 1.3
In this work, a new method is proposed to accelerate the evolution of Genetic Programming populations. The proposed method replaces the evaluation of each individual on the entire dataset by dividing the training data into small subsets and assigning each individual in the population a different subset, in order to reduce computational time. The performance of the proposed approach has been examined on 7 datasets, 4 for classification and 3 for regression, with a variety of difficulties in both groups. The results obtained demonstrate that the new approach achieves performance equivalent to the conventional algorithm of training all individuals on the entire dataset; this procedure allows for a significant reduction in computation time compared to the classical GP algorithm against which it was compared.
The Agave tequilana crop (blue Weber agave) is of immense economic relevance to Mexico, which explains the large increase in agave-dedicated land in the country. This study focuses on the application of multispectral UAV-acquired photography for precise plant recognition to enable targeted interventions and optimize resource usage in blue agave cultivation. However, the scarcity of blue agave crop image datasets and the presence of shadows in images pose challenges. To address these issues, a comprehensive dataset of 932 diverse multispectral images was created, and a novel method based on computer vision and vegetation indices, such as the atmospherically resistant vegetation index (ARVI), was proposed for effective shadow attenuation and image segmentation. The proposed approach converts images to the HSV color space and uses an inverted shadow map to make corrections in the S and V channels, resulting in enhanced ARVI calculation and improved agave plant segmentation. Experimental evaluation demonstrates notable improvements across all metrics, with enhancements of up to 10
In the context of Industry 4.0, multi-robot systems (MRS) have become essential for enhancing the adaptability and efficiency of flexible manufacturing systems, enabling rapid responses to market demands through personalized customization. Effective collaboration among multiple robots requires advanced communication, shared goal alignment, and the ability to gather and process environmental data to execute coordinated actions with precision. Despite their advantages, improving efficiency of robot learning still remains a crucial challenge, particularly in flexible manufacturing multi-robot systems, where generalization across diverse scenarios is essential for effective deployment. In this work, we propose the use of neuro-evolutionary computation to solve the particular learning-from-scratch problem in a multi-robot system. This approach consists of an architecture of a decentralized MRS in which a local controller per robot is based on Artificial Hydrocarbon Networks machine learning model. Also, it includes a learning strategy via Wound Treatment Optimization algorithm. We implement the architecture in a simulated environment to solve the mountain car domain. Experimental results and a comparison with reinforcement learning validate the ability of this approach to learn a task from scratch without prior explicit data. We anticipate the applicability of this approach in smart factories or autonomous vehicles.
Electrical Impedance Mammography (EIM) is a non-invasive method for evaluating breast tissue by generating conductivity maps across multiple depth levels. This article focuses on Gaussian Pyramid (GP) fusion, a multiscale fusion technique used to combine EIM images into a single and more representative image. By preserving both local and global features at different scales, this approach is applied to a binary classification scenario in the context of early breast cancer detection. The theoretical foundations, implementation, and results using a clinical dataset are presented. Among the classifiers evaluated through cross-validation with various k-fold configurations (k = 3, 5, 7, and 10), the SVM model achieved the best performance using 10-fold cross-validation, obtaining an accuracy of 84.21
Urban crime exhibits spatio - temporal clustering patterns known as "near repeat" phenomena. This study extends the spatial transmission network framework by He et al. (2020) to analyze day-night temporal variations in crime patterns. Using crime data from Morelia, Mexico, we introduce an eight-state taxonomy for characterizing spatial zones and transition matrices that capture temporal reconfigurations. We propose a temporal reconfiguration index (IR) to quantify urban crime pattern variability. Results demonstrate non-random crime distribution with zones acting as sources, sinks, or thoroughfares, with different crime types exhibiting varying temporal stability. Hierarchical clustering identifies distinct zone categories with specific temporal behaviors. This framework provides tools for understanding temporal crime dynamics and suggests the potential for comparative urban analysis.
Occupational health in office environments is a crucial aspect of managing the well-being and productivity of employees. However, despite the abundant research on activities in industrial, domestic, and sports settings, there is a significant gap in the recognition of activities in offices. This study aims to develop a system for analyzing ergonomic risks associated with repetitive actions and postural assessment during labor activities. The system analyzes seated posture and upper-limb activities (e.g., typing and writing) to estimate the subject’s ergonomic risk level. The proposed solution achieved an F1-score of 92.92
Classifying sensitive information is crucial for security applications like Data Leakage Prevention, but often hindered by scarce labeled data. While GAN-BERT has shown promise for low-data scenarios, its comparative efficacy against standard BERT under extended training and the impact of critical factors like unlabeled data domain alignment remain under-investigated. We empirically evaluate standard BERT fine-tuning against the semi-supervised GAN-BERT framework for this task using adapted public datasets (Monsanto, Enron) under severe low-data constraints (10 labeled examples per class). Our findings reveal that: (i) standard BERT, with sufficient training, demonstrates a powerful capacity to fit the data distribution and achieve high accuracy even with only 10 labeled examples per class, challenging the default assumption that more complex semi-supervised methods are always superior; (ii) GAN-BERT, despite faster initial convergence, plateaus earlier and incurs higher computational costs; and (iii) critically, GAN-BERT's performance is heavily dependent on the domain alignment of unlabeled data, with a smaller, aligned corpus outperforming a larger, mismatched one. This work offers a critical qualification of GAN-BERT's applicability, an in-depth analysis of performance trade-offs, and provides practical insights for applying NLP in resource-constrained, security-relevant scenarios.
This study investigates deep-learning methods for segmenting the migration of human hepatocellular carcinoma (HCC) cells in wound-healing assay images. We implemented three state-of-the-art architectures: U-Net, Attention U-Net and U-Net++, and benchmarked them with Dice score and Intersection-over-Union (IoU). Using a proprietary dataset of 177 expertly segmented images, we also performed statistical tests to quantify performance variability. Although no model out-performed the others with statistical significance, Attention U-Net achieved the highest mean scores and exhibited the most normally distributed results. For clinicians and cellular-migration analysts, these findings underscore the added value of attention mechanisms: they deliver more reliable, reproducible segmentations, thereby improving quantitative assessments of tumour-cell motility and enhancing downstream therapeutic or diagnostic research.
