Carlos Castorena

About me

Mechatronics Engineer with a Master's degree in Engineering Sciences, awarded by the Instituto Tecnológico de Toluca in 2017 and 2021 respectively. Currently pursuing a Ph.D. in Informatics, Communications, and Computer Technologies at the University of Valencia, under a predoctorar contract Santiago Grisolia, focusing on the field of artificial intelligence, with a specific emphasis on intelligent audio processing. I had the privilege of initiating the academic career by participating as an author in various articles and presentations at national and international conferences. These initial achievements reflect a commitment to advancing research and enthusiasm for contributing to knowledge in the field.

Work experience

2021-Currently

Predoctoral researcher. University of Valencia

2021-2021

Artificial intelligence expert. Centro de Experimentación y Seguridad Vial (Cesvi)

During my seven-month work at the Center for Experimentation and Road Safety (Cesvi) in Mexico, from February 22 to September 24, 2021, I held a full-time position in the Research Department under a permanent contract. In this role, I specialized in designing and implementing artificial intelligence algorithms for diverse projects aligned with the company's objectives, contributing to my growth in this dynamic field.

2019-2021

Proffesor, Instituto Tecnológico de Toluca

2015-2019

Metrology Engineer and Quality Laboratory Manager. Valeo Sistemas Electricos

During my four-year tenure at Valeo Sistemas Electricos in Toluca, Mexico, from February 23, 2015, to January 26, 2019, I held a full-time position in the Quality Department under an indefinite employment contract. Serving as both a Metrology Engineer and Quality Laboratory Manager, I played a crucial role in ensuring precision and excellence in quality control processes. This experience honed my skills in metrology and leadership, contributing significantly to my professional journey.

My papers

A safety-oriented framework for sound event detection in driving scenarios (2023)

C Castorena, M Cobos, J Lopez-Ballester, FJ Ferri. Applied Acoustics 215, 109719

The safety of drivers has become increasingly important in today's rapidly evolving transportation landscape, especially with the rise of autonomous and smart vehicles. This paper proposes a safety-oriented framework for sound event detection in smart vehicles using deep learning models. The goal of this framework is to increase driver awareness, prevent accidents, and provide acoustic forensic analysis. To achieve this, a meaningful taxonomy of event classes in a driving scenario is introduced, taking into account the event classes that are known to be related to major driving distractors. Based on this taxonomy, a dataset has been created to train and evaluate a fully-convolutional sound event detection model that was inspired by the well-known YOLO vision model. Experimental results demonstrated that the proposed model offers competitive results, outperforming a state-of-the-art baseline using recurrent connections. This comprehensive framework for sound event detection in smart vehicles aligns with the recommended directions for future mobility scenarios and has the potential to significantly improve the safety and performance of smart vehicles.

Mejoras En La Detección De Eventos Acústicos Mediante La Ponderación De Desequilibrio De Actividad Y Predicciones Débiles Basadas En El Máximo (2023)

C Castorena, FJ Ferri, M Cobos. Congreso Español de Acústica (TECNIACÚSTICA)

La detección de eventos de sonido es uno de los temas más relevantes en el procesamiento de audio con redes neuronales artificiales, en este trabajo, se presentan dos estrategias para mejorar la detección de eventos de audio en el sistema base de la Tarea 4 del DCASE 2022. La primera estrategia propuesta se enfoca en mejorar la generación de predicciones débiles a partir de las predicciones de actividad del sistema, esto resulta en un cálculo de pérdida más sólido y coherente para ejemplos con etiquetas débiles, lo que a su vez mejora la capacidad del sistema para identificar eventos de audio con mayor precisión. La segunda estrategia se basa en una pérdida ponderada de entropía cruzada binaria, que tiene en cuenta el desequilibrio resultante de la actividad de eventos en cada lote de entrenamiento. Esto garantiza que el sistema se entrene de manera más equilibrada y pueda capturar de manera efectiva los eventos de audio en diferentes proporciones. La combinación de ambas estrategias ha demostrado resultados prometedores, ya que se lograron mejoras en el rendimiento del sistema sin requerir un esfuerzo considerable en el diseño y entrenamiento del mismo.

Comparing the Use of Soft and Hard Labels for Sound Event Detection (2022)

C Castorena, FJ Ferri, M Cobos. Union Radio-Cientifique International (URSI)

Sound event detection with deep learning models such as convolutional recurrent networks have shown highly competitive results in performing this task. However, as they are supervised algorithms, the results are directly affected by the quality of the labeling, especially for events with an intermittent time activation. In this work, we evaluate the detection performance of a baseline architecture trained on artificial mixtures of events using the energy envelopes of isolated events as soft labels. Unlike with hard binary labeling, this allows not only to know the presence of the event, but also provides a prediction of its energy level, leading to a more informative description of the event temporal activation. The experiments show that the use of non-binary labels generate highly competitive results with respect to the state of the art and are a viable option to perform this task in the context of intermittent sounds.

Artificial Neural Networks for COVID-19 Forecasting in Mexico: An Empirical Study (2022)

CM Castorena, R Alejo, E Rendón, EE Granda-Gutíerrez, RM Valdovinos. International Conference on Intelligent Computing, 168-179.

Artificial Neural Networks (ANN) have encountered interesting applications in forecasting several phenomena, and they have recently been applied in understanding the evolution of the novel coronavirus COVID-19 epidemic. Alone or together with other mathematical, dynamical, and statistical methods, ANN help to predict or model the transmission behavior at a global or regional level, thus providing valuable information for decision-makers. In this research, four typical ANN have been used to analyze the historical evolution of COVID-19 infections in Mexico: Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), Long Short-Term Memory (LTSM) neural networks, and the hybrid approach LTSM-CNN. From the open-source data of the Resource Center at the John Hopkins University of Medicine, a comparison of the overall qualitative fitting behavior and the analysis of quantitative metrics were performed. Our investigation shows that LSTM-CNN achieves the best qualitative performance; however, the CNN model reports the best quantitative metrics achieving better results in terms of the Mean Squared Error and Mean Absolute Error. The latter indicates that the long-term learning of the hybrid LSTM-CNN method is not necessarily a critical aspect to forecast COVID-19 cases as the relevant information obtained from the features of data by the classical MLP or CNN.

On the Performance of Deep Learning Models for Respiratory Sound Classification Trained on Unbalanced Data (2022)

C Castorena, FJ Ferri, M Cobos. Iberian Conference on Pattern Recognition and Image Analysis, 143-155.

The detection of abnormal breath sounds with a stethoscope is important for diagnosing respiratory diseases and providing first aid. However, accurate interpretation of breath sounds requires a great deal of experience on the part of the clinician. In the past few years, a number of deep learning models have been proposed to automate lung classification tasks in physical examination. Unfortunately, acquiring accurately annotated data for this problem is not straightforward and important issues arise, as the available examples of abnormal and normal sounds usually differ substantially. This work provides a comprehensive analysis of deep learning models making use of different class balancing methods during training, considering multiple network architectures and audio input features. The results show that good performance is achievable when applying random oversampling and a convolutional neural network operating over Mel-frequency cepstral coefficient (MFCC) representations.

Deep neural network for gender-based violence detection on Twitter messages (2021)

CM Castorena, IM Abundez, R Alejo, EE Granda-Gutiérrez, E Rendón. Mathematics 9 (8), 807.

The problem of gender-based violence in Mexico has been increased considerably. Many social associations and governmental institutions have addressed this problem in different ways. In the context of computer science, some effort has been developed to deal with this problem through the use of machine learning approaches to strengthen the strategic decision making. In this work, a deep learning neural network application to identify gender-based violence on Twitter messages is presented. A total of 1,857,450 messages (generated in Mexico) were downloaded from Twitter: 61,604 of them were manually tagged by human volunteers as negative, positive or neutral messages, to serve as training and test data sets. Results presented in this paper show the effectiveness of deep neural network (about 80% of the area under the receiver operating characteristic) in detection of gender violence on Twitter messages. The main contribution of this investigation is that the data set was minimally pre-processed (as a difference versus most state-of-the-art approaches). Thus, the original messages were converted into a numerical vector in accordance to the frequency of word’s appearance and only adverbs, conjunctions and prepositions were deleted (which occur very frequently in text and we think that these words do not contribute to discriminatory messages on Twitter). Finally, this work contributes to dealing with gender violence in Mexico, which is an issue that needs to be faced immediately.

Deep Neural Network to Detect Gender Violence on Mexican Tweets (2021)

G Miranda, R Alejo, C Castorena, E Rendón, J Illescas, V García. Progress in Artificial Intelligence and Pattern Recognition.

During COVID-19 quarantine, in online sites such as social networks, Gender-Based Violence has alarmingly increased. Online platforms have taken various measures to regulate and prevent broadcasting of violence messages. Multiple proposals based on machine learning and deep learning approaches have been used to address this problem. This work presents an improvement in implementation of a deep learning neural network for detection of Gender-Based Violence in Twitter messages. A total of 32,500 tweets were downloaded from Mexican Twitter accounts and human volunteers manually tagged the tweets as violent and non-violent to be used as training and testing data sets. Experimental results show the effectiveness of the deep neural network (about 90% of the Area Under the Receiver Operating Characteristic) to detect gender violence in Twitter messages using a simple Natural Language Processing approach.

Comparative study of methods to obtain the number of hidden neurons of an auto-encoder in a high-dimensionality context (2020)

HR Vega-Gutierrez, C Castorena, R Alejo, EE Granda-Gutierrez. IEEE Latin America Transactions 18 (12), 2196-2203.

Fourteen formulas from the state-of-art were used in this paper to find the optimal number of neurons in the hidden layer of an autoencoder neural network. The latter is employed to reduce the dataset dimension on high-dimensionality scenarios with not significant reduction in classification accuracy in comparison to the use of the whole dataset. A Deep Learning neural network was employed to analyze the effectiveness of the studied formulas in classification terms (accuracy). Eight high-dimensional datasets were processed in an experimental set in order to assess this proposal. Results presented in this work show that formula 13 (used to find the number of hidden neurons of the auto-encoder) is effective to reduce the data dimensionality without a statistically significant reduction of the classification performance, as it was confirmed by the Freidman test and the Holm's post-hoc test.

Data sampling methods to deal with the big data multi-class imbalance problem (2020)

E Rendon, R Alejo, C Castorena, FJ Isidro-Ortega, EE Granda-Gutierrez. Applied Sciences 10 (4), 1276.

The class imbalance problem has been a hot topic in the machine learning community in recent years. Nowadays, in the time of big data and deep learning, this problem remains in force. Much work has been performed to deal to the class imbalance problem, the random sampling methods (over and under sampling) being the most widely employed approaches. Moreover, sophisticated sampling methods have been developed, including the Synthetic Minority Over-sampling Technique (SMOTE), and also they have been combined with cleaning techniques such as Editing Nearest Neighbor or Tomek’s Links (SMOTE+ENN and SMOTE+TL, respectively). In the big data context, it is noticeable that the class imbalance problem has been addressed by adaptation of traditional techniques, relatively ignoring intelligent approaches. Thus, the capabilities and possibilities of heuristic sampling methods on deep learning neural networks in big data domain are analyzed in this work, and the cleaning strategies are particularly analyzed. This study is developed on big data, multi-class imbalanced datasets obtained from hyper-spectral remote sensing images. The effectiveness of a hybrid approach on these datasets is analyzed, in which the dataset is cleaned by SMOTE followed by the training of an Artificial Neural Network (ANN) with those data, while the neural network output noise is processed with ENN to eliminate output noise; after that, the ANN is trained again with the resultant dataset. Obtained results suggest that best classification outcome is achieved when the cleaning strategies are applied on an ANN output instead of input feature space only. Consequently, the need to consider the classifier’s nature when the classical class imbalance approaches are adapted in deep learning and big data scenarios is clear.

Funded R&D&i projects

PID2022-137048OB-C4. SpaTial Audio and ARRay processing for INdustrial Applications and diGital Transformation (STARRING) TED2021-131003B-C21. Monitorización acústica y multimodal asistida por inteligencia artificial: computación afectiva y reconocimiento de eventos para la seguridad en la conducción (AEOLIAN). AICO/2022/057. Aprendizaje Automático aplicado a Individualización de SONido Espacial InMersivo y Síntesis de Espacios Virtuales 3D (SONEM3D) RTI2018-097045-B-C21. Audio Espacial Inteligente: Análisis y Monitorización (ISLA-ALAMO). AICO/2020/154. Técnicas de aprendizaje automático aplicadas al análisis computacional de escenas sonoras y la síntesis de entornos inmersivos.

Key Courses and Formation

Postgraduate in Master of Engineering Sciences, México Degree in Mechatronics Engineering, México

Other mentions

Scholarship: Santiago Grisolía de la Consellería de Innovación, Universidades, Ciencia y Sociedad Digital, Spain Research stay at the Jaume Primer University, Castellon de la Plana, Spain. Conacyt national postgraduate scholarship, México First place in the presentation and poster competition, presentation category, of the CIIA 2019 International Congress of Administrative Informatics.