Poster Sessions 2025

Adversarial perturbations are specially crafted noise patterns designed to degrade the performance of machine learning models. In this study, we investigate their impact on deep learning-based ultrasound beamformers. We evaluate three previously reported end-to-end DNN-based beamformers trained on simulated ultrasound data. These models are assessed under both adversarial perturbations and realistic distortions. Our results reveal a correlation between the distortions caused by adversarial perturbations and those observed in realistic scenarios— a relationship not evident with Gaussian or uniform noise. These findings provide insights into the robustness and generalization of DNN-based beamformers, particularly in scenarios where simulated data is used to train deep learning models, as is common in medical imaging applications.

The development of Large Language Models (LLMs) has transformed natural language processing, yet their reliance on extensive computational resources poses challenges for regions with limited infrastructure, such as Brazil. This work evaluates language models, including general-purpose models and Portuguese-specialized alternatives, across question answering, text simplification, and summarization tasks in Portuguese. The assessment considers both generative quality and computational efficiency, addressing sustainability concerns. Results highlight the importance of adapting models to linguistic and cultural specificities. By analyzing performance across diverse tasks and metrics, this work aims to contribute to democratizing LLM development and application in low-resource settings, fostering research aligned with regional realities and sustainability goals.

Social network users commonly propose new expressions to make communication faster, more fluid, and more original. Moreover, although they follow a specific idiom, the employed language is also mixed-coded, making even the same idiom go beyond the usual variations. This way, texts in Portuguese from different regions that already have linguistic, syntactic, lexical, and orthographic differences go to a new level when written in social media posts. This work proposes relying on Portuguese language models to analyze the variations of Portuguese in social media posts in Brazil, Portugal, and Mozambique.

This poster presents a curriculum proposal and didactic materials designed to promote a critical approach to teaching AI in schools. The AI Curriculum proposal is part of Argentina’s broader initiative to include Computer Science in compulsory education. Developed by the Program.AR Initiative of the Sadosky Foundation and researchers from various universities, this proposal defines key AI and CS concepts necessary for full citizenship in a technology-driven society. It outlines a progressive learning path from early childhood to high school, emphasizing data analysis, AI applications, and ethical considerations. Additionally, we share findings from the implementation of an AI workshop for high school students and teacher training programs. The workshop integrates technical and ethical perspectives, focusing on biases in AI models. Participants engage in hands-on activities, experiencing how decisions in AI development introduce biases. Pre and post-test analysis revealed a significant shift in participants’ perceptions, reducing beliefs in AI’s neutrality and superiority over human decision-making. The workshop was perceived as engaging and practical, leading to the creation of educational materials for broader classroom implementation. This work contributes to ongoing discussions on AI education, offering tools for critically analyzing and understanding the role of AI in contemporary society.

Not available yet

The intersection of **Machine Learning (ML) and Causal Inference** holds great promise for advancing healthcare by enabling robust decision-making and personalized treatment strategies. Our research group focuses on **developing and applying causal ML methodologies** to tackle key challenges in **mental health, ECG analysis, and diabetic retinopathy diagnosis**. In mental health, we leverage **interventional modeling** to assess the impact of treatment strategies on patient outcomes, aiding in the design of personalized interventions. For ECG analysis, we integrate ML with **counterfactual prediction frameworks** to estimate the potential effects of different clinical interventions on cardiac health, addressing treatment-effect heterogeneity. In the context of diabetic retinopathy diagnosis, we employ **causal ML techniques** to mitigate confounding biases in predictive models, improving the reliability of automated screening tools. By bridging **causal inference and ML**, our work aims to move beyond correlation-based approaches, enabling more **trustworthy, interpretable, and actionable** healthcare predictions. Our research contributes to the growing field of **causal ML for healthcare**, fostering collaboration between machine learning practitioners and medical experts to improve patient outcomes.

Automatic symptom detection from electronic health records is a valuable source for event-based surveillance systems. In this study, we develop tools to automatically detect symptoms associated with febrile illnesses in electronic health records written in Spanish. Therefore, we use a custom corpus, comprising 6228 expertly labeled and approximately 1 million unlabeled health reports. Our approach involved fine-tuning state-of-the-art named entity recognition models, including BiLSTM-CRF and transformer-based models like RoBERTa. We focused on domain-adaptive and task-adaptive models to enhance performance: the former were pre-trained on biomedical corpora, while the latter were further pre-trained on our unlabeled health reports. Despite computational constraints, our models demonstrated promising results, with RoBERTa-Clinico, a task-adaptive transformer model pre-trained in our unlabeled corpus, showing the best micro recall performance (79.30), and 70.83 micro F1 score, which are comparable to results in similar studies. In this way, we contribute to the limited body of work in BioNLP in Spanish.

Online platforms facilitate the exchange of opinions but also enable the spread of hate speech, including antisemitism. Despite its growing presence in digital news comments, most research on hate speech detection focuses on English, leaving Spanish-language content underexplored.This study investigates antisemitic discourse in digital newspaper comments from Argentina, Chile, Colombia, Costa Rica, Panama, and Uruguay. It identifies temporal and spatial patterns and examines distinctive linguistic features of antisemitic comments. A manually annotated corpus of 13,511 comments from 2,834 news articles mentioning "Israel" since February 2021 was compiled. The study aims to present and describe this dataset while analyzing its lexical characteristics.Unigram analysis identifies characteristic terms in antisemitic versus non-antisemitic comments. Additionally, the study develops automatic detection tools using large language models, such as GPT-4 and Llama 3.2 3B, implementing few-shot and zero-shot learning strategies through prompt engineering. Model performance is evaluated with metrics that account for class imbalance, including the F1 score.This research advances antisemitic discourse detection in Spanish and contributes to the development of essential tools to address the growing phenomenon of online hate speech targeting cultural and religious minorities.

This work presents an automatic Tree Ring Delineation method for complete cross-sections of conifer tree images (CS-TRD) and its extension to angiosperms and shrubs (DeepCS-TRD). The method detects, processes, and connects edges corresponding to tree growth rings. It relies on edge detection parameters and the precise location of the pith—the soft central region of the wood slice, which is concentric to all rings but not always positioned at the geometric center.The pith location is estimated using an automated method based on the concentric structure of tree rings (APD). This approach leverages local ring orientations, estimated via the 2D structure tensor, to identify the optimal pith candidate through a dedicated cost function.DeepCS-TRD enhances CS-TRD by replacing the edge detection step with a deep-learning-based segmentation model (U-Net), enabling its application across diverse imaging modalities—including microscopy, scanner, and smartphone images—and various species.A graphical user interface was developed to compute key wood cross-section metrics such as growth area and perimeter. Users can manually refine ring detections through an intuitive interface.

Understanding how audio models make predictions is crucial for their interpretability and trustworthiness. In this work, we explore temporal explainability in audio models, aiming to generate explanations that evolve over time and align with the model’s decision-making process. We focus on developing robust evaluation methods to assess the quality of these explanations, ensuring they provide meaningful insights. Additionally, we leverage these explanations to uncover biases and hidden correlations in audio representations

Motivated by applications of large embedding models, we study differentially private (DP) optimization problems under sparsity of individual gradients. We start with new near-optimal bounds for the classic mean estimation problem but with sparse data, improving upon existing algorithms particularly for the high-dimensional regime. Building on this, we obtain pure- and approximate-DP algorithms with almost optimal rates for stochastic convex optimization with sparse gradients; the former represents the first nearly dimension-independent rates for this problem. Finally, we study the approximation of stationary points for the empirical loss in approximate-DP optimization and obtain rates that depend on sparsity instead of dimension, modulo polylogarithmic factors.

Este estudio presenta un método para identificar textos relacionados con el cambio climático y la sustentabilidad en español mediante modelos de lenguaje preentrenados. El objetivo es desarrollar una herramienta que identifique, entre distintos textos, aquellos que aborden temas como el cambio climático y sus efectos. Se creó un dataset etiquetado con textos en español, se entrenó el modelo con estos datos y se evaluó utilizando un conjunto de validación, logrando una precisión de 0.916. F1 Score de 0.95 y Recall de 0.99. Los resultados muestran la efectividad del enfoque propuesto para identificar textos relevantes sobre el cambio climático en español. Este modelo puede ser una herramienta útil para crear y enriquecer repositorios sobre el tema, contribuyendo a una mayor conciencia y comprensión sobre el cambio climático y sus implicaciones.

This study examines the implementation of generative artificial intelligence (GenAI) to optimize data governance in a Chilean social security entity. Facing current challenges in personal data protection and regulatory compliance, a solution based on fine-tuning and prompt engineering was developed for Google's Gemini foundation model, which automates key metadata management processes. The system implements a pipeline of five main tasks: generation of field and table descriptions, identification of personal and sensitive data, and classification according to business domains. The evaluation through quantitative and qualitative metrics demonstrated high precision in sensitive data identification and significant improvements in operational efficiency. Despite identified technical and organizational limitations, the results validate the feasibility of digital transformation in data governance through GenAI, establishing a precedent for its application in social security organizations.

The landscape of the natural sciences, particularly within computer science, is evolving at an increasingly rapid pace. Scholars face growing challenges in analyzing and synthesizing emerging research before it becomes outdated. Large Language Models (LLMs) currently demonstrate superhuman capabilities in processing and analyzing information at remarkable speeds; however, they are susceptible to generating hallucinations when queried for specific information beyond the scope of their training corpus. In this context, we introduce Knowledge Retrieval and Analysis for Scientific Support (KRASS), an open-source algorithm designed to facilitate literature reviews and state of the art follow-up on a given subject. KRASS leverages Large Language Models within a Conditional Retrieval-Augmented Generation (CRAG) framework to evaluate, retrieve, refine, and analyze academic literature in response to specific expert queries. We evaluated the performance of our algorithm across various Large Language Models and academic topics, and assessed its ability to retrieve sources and generate summaries from the selected literature in response to user queries.

Visual processing in the human cerebral cortex is hierarchical and highly parallelized. Similarly, in the deep learning architecture known as vision transformers, data is processed iteratively through attention operations, which allow for the parallel evaluation of the importance of certain aspects of the image in the context of the rest of the image. In both the human brain and transformers, the attention mechanism selects parts of the input that are more relevant to the task at hand, enabling the model to prioritize and weigh this information during processing. In this work, we investigate the temporal correspondence between both systems, based on the hypothesis that each successive layer of the transformer will be correlated with temporally ordered components of the EEG response in an object recognition task. First, we analyze the representational similarity between the two systems, then we investigate predicting brain activations based on neural network activations, and finally, we use a state-of-the-art metric to make comparisons between neural network representations. We find that the early layers of the network are mostly aligned with the early stages of electrophysiological signals, while the advanced layers correspond to later stages.

Handling conflicting objective functions in machine learning, such as those related to fairness, class imbalance, reinforcement learning, multi-task learning, or multi-class classification, is not an easy task. Using a single loss function, it is hard to properly optimize these conflicting objectives of reducing false positive and false negative rates simultaneously. For that reason, multi-objective optimization methods can be very advantageous as they minimize incorrect answers by tuning the machine to automatically promote lower false positive or false negative rates with different trade-offs, resulting in a set of efficient classifiers.While there are different methods to deal with this challenge, a comprehensive framework for multi-objective optimization in machine learning contexts remains lacking. This research aims to fill this gap by formalizing and implementing such a framework, along with exploring and contextualizing multi-objective optimization methods and their application in machine learning, especially in fair models. Additionally, the goal is to offer a user-friendly implementation through a Python package that can be integrated with other machine learning libraries, such as Scikit-learn, TensorFlow, and PyTorch, to facilitate its use in machine learning problems.

This work studies the feasibility of automatically detecting consultations on sexual and reproductive health (SSH) in the electronic health records (EHR) of people with disabilities (PWD) with machine learning models. This could be an input to improve access to SSH for this population through the adaptation of future consultations, obtaining better statistics, and supporting health management planning.An interdisciplinary team worked on: the construction of operational definitions of SSH and PWD consultation, the identification of entities in free and structured text of the EHR, the development of a manually annotated data set for training the models, obtaining the classification models for automatic detection and their evaluation with equity metrics.First, a classifier for identifying SSH queries and another for identifying if the query corresponds to PWD where created. Then, the concatenation of these classifiers was considered as the final model.As a result, the accuracy of these models is ~80% for each of these categories separately. However, for the simultaneous classification of SSH and PWD, the accuracy is 62%, which is much higher than that based on regular expressions and logical rules (with an accuracy of 19%). This allows us to guide future work in improving the proposed models.

This study is focused on developing an artificial intelligence (AI) model that predicts increases in river flow fluctuations to improve hydro-climatic risk management, thereby mitigating economic losses and saving lives. The research utilizes 26,788 hydrometeorological data points from Colombia (1950–2024) along with climate indices provided by the National Oceanic and Atmospheric Administration (NOAA). The key inputs for the model are the river flow values and corresponding dates, which are merged with the time series of the climatical indices obtaining a multivariate time series.Initially, the model forecasts a single value; however, it is designed to generate multi-step predictions using neural networks. A Long Short-Term Memory (LSTM) model is proposed, leveraging its capability to handle long-term dependencies in multivariate time series. The model is evaluated primarily using the Root Mean Squared Error (RMSE), and its precision and robustness are further validated with complementary metrics: MSE, MAE and RMSLE. Alternative approaches such as Recurrent Neural Networks (RNN) and statistical models (ARIMA, SARIMA) were initially considered, but these exhibit certain limitations. The state-of-the-art review highlights previous models—like cascade feedforward networks and ARIMA—that face constraints in terms of temporal dynamics and variable handling.

The evaluation of vision large language models (VLLMs) has mainly relied on English-language benchmarks, leaving significant gaps in both multilingual and multicultural coverage. While multilingual benchmarks have expanded, many rely on translations of English datasets, failing to capture cultural nuances. In this work, we propose Global Multimodal and Multilingual Exams (Global MuMu Exams) as the most comprehensive exams benchmark for multilingual evaluation to-date of vision models. Global MuMu Exams is a large-scale, in-language multimodal benchmark designed to evaluate MLLMs across diverse languages and visual inputs. We cover 18 languages, 14 different subjects, and a total of 20,911 multiple-choice questions, of which 55% require image understanding. Built through a direct collaboration with a diverse group of researchers worldwide, Global MuMu Exams ensures linguistic and cultural authenticity. We evaluate top performing multilingual vision-language models, and find that they perform poorly on low-resource languages and in complex multimodal scenarios. Our results highlight the need for progress on culturally inclusive multimodal evaluation frameworks.

We'll present partial results of our ongoing research (in partnership with DASA) in tumor detection and classification in breast MRI's. Mainly focusing on the results obtained using CNN's and the impacts of data augmentation. Although our objectives are detection and classification, we'll present only detection, since classification is still in early development.

Functional Harmony, or Roman Numeral Analysis, studies the roles of chords within a musical progression, categorizing them into functions like tonic (stability), subdominant (instability and change), and dominant (dissonance and transition). This task is challenging because a chord's function depends on the overall musical context, and the same chord can have different roles even within the same piece. While deep learning has advanced symbolic music analysis, directly analyzing functional harmony from audio signals remains a complex and underexplored problem. We propose DeepToneNets (DTNs), a novel deep-learning approach that combines neural networks for harmonic and bass chromagram extraction with the Tonnetz, a geometric representation of tonal relationships. We evaluate the impact of the Tonnetz and compare DTNs to state-of-the-art methods in both symbolic and audio domains. Our initial results demonstrate promising trends, with room for improvement. We also highlight the potential of integrating audio-based harmonic analysis with deep learning, providing a foundation for future research.

While research shows that people can be vulnerable to adversarial perturbations, it remains unclear whether their influence operates similarly to artificial neural networks. Here, we investigate whether biases arise from explicit semantic content or if subtle attacks distort perception, causing subjects to exhibit bias without confidently articulating their reasoning.To address this, we conducted two forced-choice experiments with 20 participants: one with brief stimulus exposure and another with unlimited exposure. Adversarial perturbations of different magnitudes were designed to mislead an ensemble of six convolutional neural networks. Human responses, confidence levels, reaction times, and eye movements were recorded, and a post-hoc survey was conducted in the unlimited exposure experiment.Results indicate that in brief exposure, participants predominantly selected the original category, and when they did choose the adversarial category, they did so with low confidence. In one category pair, statistical significance showed that adversarial perturbations successfully misled participants. However, in unlimited exposure, perturbations had an evident effect, shown by increased confidence when selecting the adversarial image. This effect was linked to participants noticing alterations meant to deceive their classification or identifying semantic features resembling the target category, implying that misleads occur when adversarial attacks are evident rather than subtle.

Oceans, covering over 70% of Earth’s surface, regulate the global climate by absorbing heat and CO2. Marine microbiomes support food webs and drive the biological carbon pump. However, the relationships between taxonomy, gene expression, metabolism, and environmental factors remain poorly understood.We applied Explainable Artificial Intelligence using Symbolic Regression and a SHAP-based approach to model these complex interactions with the OMR Gene Catalog, which contains 47 million genes annotated across 9,024 molecular functions. We generated 36,214 interpretable models, enabling the prediction of: (i) omics-derived features from environmental factors, (ii) relationships within taxonomic compositions, and (iii) taxonomic composition based on functional traits.Our results revealed that environmental factors predicted 22.6% and 29.4% of molecular functions and metabolic pathways from metagenomes, while only 4.9% and 3.5% were predicted from metatranscriptomes (R2 ≥ 0.5). The most accurate predictions involved molecular functions linked to carbon sequestration pathways. By providing a dataset of equations predicting omics-derived features, our work enables data-driven hypotheses on microbial function and taxonomic dynamics under environmental changes, as well as simulations of potential shifts under climate change scenarios.

The SOPHIE spectrograph, operational since 2006, enables detection of Earth-mass exoplanets through high-precision radial velocity (RV) measurements. However, systematic effects can cause variations in the nightly zero point (NZP), which are traditionally corrected using regular observations of standard stars. These systematic variations are believed to stem from residual environmental changes within the instrument and vary depending on the characteristics of the target star. We present a novel data-driven approach using machine learning to model and correct these NZP variations, incorporating housekeeping variables, stellar properties, and observational conditions. To validate our method, we tested it on 12 SOPHIE targets with known planetary companions, analyzing each system using three different sets of RVs: those processed with our correction method, those using the traditional master correction, and uncorrected data. Results demonstrate that both correction methods consistently outperformed uncorrected data, with our approach showing superior results in approximately half the cases. This demonstrates the effectiveness of our machine learning-based approach as a viable alternative to traditional NZP correction methods in high-precision RV measurements.

Machine learning models, particularly sub-symbolic ones, have experienced rapid development and adoption across various fields. However, their complexity and opacity pose significant challenges for interpretation and justification. The field of Formal Explainable AI (FXAI) has emerged to provide rigorous, theoretically grounded explanations, addressing the limitations of heuristic explainability methods.Most work in formal explainability has focused on Boolean models. However, to address the inherent uncertainty in many real-world problems, it is necessary to consider models that incorporate probabilistic aspects. In this context, probabilistic circuits (PC) present themselves as a particularly promising class of probabilistic graphical models (PGM), combining the expressivity of neural networks with formal properties that facilitate their analysis.Our main objective is to develop the theoretical and practical foundations of a declarative language that allows users to formulate and execute explainability queries in PC models. Based on the extension of First-Order Interpretability Logic, we analyze expressivity limits, design tractable extensions, study computational complexity, and develop corresponding algorithms. This work contributes to FXAI by proposing an approach to improve the explainability of PGM.

Machine Learning is a fundamental tool for information analysis. Among its various techniques, clustering stands out as a family of algorithms capable of dividing large datasets into distinct groups based on similarity. In the healthcare domain, state-of-the-art research has been conducted, leveraging the vast availability of patient medical data, which makes clustering a powerful tool for knowledge discovery. However, Machine Learning also presents limitations, such as difficulties in explaining its results and the potential for unethical biases, which pose significant challenges for real-world applications. This study explores opportunities for applying clustering techniques within the Social Security Insurance system of Universidad Nacional del Comahue, a university located in Neuquén, Argentina. Additionally, clustering will be combined with the SHAP framework to enhance the explainability of the obtained results.

Identifying propaganda in social media posts is essential for understanding the strategies employed by policymakers and stakeholders to shape public perception. In this poster, we present our approach to the automated detection and characterization of propaganda techniques and narratives used by diplomats from major powers on social media. Our primary focus was to address class imbalances in the datasets provided for Task 1 of the DIPROMATS 2024 competition.Our initial evaluation revealed that certain propaganda techniques were underrepresented in the dataset. To address this, we implemented two data augmentation strategies to improve performance in these categories. The first strategy involved translating text samples between Spanish and English, enriching the dataset with new instances in both languages. The second strategy leveraged large language models (LLMs) to generate paraphrased samples, facilitating cross-lingual transfer between Spanish and English.Our solution ranked first in Subtask-1a for English (ICM score: 0.2123) and in the bilingual evaluation (ICM score: 0.2048).

Human visual search is a complex, multifaceted process involving a sequence of eye movements to locate a target. While deep neural networks can predict eye movement patterns, they offer limited insight into the underlying cognitive mechanisms of attention, decision making, and working memory. By contrast, white-box approaches that explicitly model separate cognitive processes can clarify how the brain orchestrates search and guide AI systems under real-life constraints. We propose an iterative Bayesian decision-making model leveraging both prior knowledge and new information from each fixation. The prior is based on a saliency map while each fixation adds new information about the image. We are extending our model with segmentation and large language models to incorporate contextual information. By dissecting visual search into interpretable components, our framework offers a transparent way to study its cognitive foundations in humans and machines. Finally, our model closely replicates human behaviour across the datasets in the public ViSioNS benchmark.

In this paper, we introduce the Generalized Graph Variational Auto-Encoder (GGVA), a novel framework that extends traditional Variational Graph Autoencoders by incorporating generalized divergence measures for distribution regularization in the latent space. Our main contributions are twofold: (1) a flexible variational inference framework for graph-structured data using generalized divergence measures, and (2) the empirical demonstration that these measures can improve accuracy in handling complex network topologies for link prediction tasks.

This study evaluates the recently proposed Document Attention Network (DAN) for extracting key-value information from Uruguayan birth certificates, handwritten in Spanish. We investigate two annotation strategies for automatically transcribing handwritten documents, fine-tuning DAN with minimal training data and annotation effort. Experiments were conducted on two datasets containing the same images (201 scans of birth certificates written by more than 15 different writers) but with different annotation methods. Our findings indicate that normalized annotation is more effective for fields that can be standardized, such as dates and places of birth, whereas diplomatic annotation performs much better for fields containing names and surnames, which can not be standardized.

Chile’s mining industry, a global leader in copper production, faces challenges due to the increasing volumes of mining waste. These materials are disposed of in Mine Waste Storage Facilities (MWSFs), which pose significant environmental risks due to their large volumes and hazardous contents. These include Waste Rock Dumps (WRDs), Leaching Waste Dumps (LWDs), and Tailings Storage Facilities (TSFs). The National Geology and Mining Service (SERNAGEOMIN) requires the assessment of the physical stability (PS) of these facilities, but current methods are hindered by data scarcity and resource constraints. Our recent work proposed a simplified assessment methodology using first-order parameters derived from open-access data, integrating Sentinel-2 satellite imagery, geographic information systems (GIS), and artificial intelligence (AI) to automate the detection, segmentation, and remote evaluation of PS for WRDs and LWDs in the Antofagasta region. This study preliminarily extends the methodology to TSFs, incorporating symmetry properties into AI models to enhance the characterization of mining deposits despite satellite image limitations and their diverse visual representations.

This paper presents the first comprehensive interpretability analysis of a Transformer-based Sign Language Translation (SLT) model, focusing on the translation from video-based Greek Sign Language to glosses and text. Leveraging the Greek Sign Language Dataset, we examine the attention mechanisms within the model to understand how it processes and aligns visual input with sequential glosses. Our analysis reveals that the model pays attention to clusters of frames rather than individual ones, with a diagonal alignment pattern emerging between poses and glosses, which becomes less distinct as the number of glosses increases. We also explore the relative contributions of cross-attention and self-attention at each decoding step, finding that the model initially relies on video frames but shifts its focus to previously predicted tokens as the translation progresses. This work contributes to a deeper understanding of SLT models, paving the way for the development of more transparent and reliable translation systems essential for real-world applications.

Segmentation of lung regions affected by COVID-19 in chest CT scans is vital for reliable diagnosis and effectiveevaluation of the disease’s prognosis. Few studies establish a faircomparison of the segmentation methods of CT scans for auto-mated COVID-19 detection using datasets from the literature.Thus, hindering the ability to establish certainty in identifyingwhich method demonstrates consistency. Existing methods oftenrely on independent evaluations using diverse image datasets,leading to inconsistencies and limited comparability. This studyaims to set up a single experimental environment to ensure a fairassessment of the performance of methods from the literature.Additionally, this study incorporates an interpretability analysisbased on the Explainable artificial intelligence discipline, arecent research direction that aims to interpret deep learningmodels’ results. The research framework involves two stages:First, deep learning models are evaluated using standard methods such as dice score, pixel-wise precision, and intersectionover union metrics. Second, the application of the ConfusionMatrix Heatmaps technique to interpret model predictions byvisualizing the feature detection process, intending to benchmark their performance and understand their ability to detect and recognize critical features in medicalimaging tasks.

We explored the application of artificial intelligence to enhance our understanding of embryonic development. Previously, we introduced EmbryoNet, a deep convolutional neural network crafted for the automated phenotyping of zebrafish signaling mutants. This tool effectively identifies and classifies phenotypic anomalies induced by disruptions in critical signaling pathways. Currenlty, we are exploring the temporal evolution of embryo development using recurrent neural networks and autoencoders, aiming to uncover dynamic patterns and temporal relationships that drive developmental processes. These methodologies not only standardize phenotypic classification but also facilitate high-throughput drug screening and comparative embryology across various species.

The use of Large Language Models (LLMs) has proven to be a tool that could help in the automatic detection of sexism. Previous studies have shown that these models contain biases that do not accurately reflect reality, especially for minority groups. Despite various efforts to improve the detection of sexist content, this task remains a significant challenge due to its subjective nature and the biases present in automated models.We explore the capabilities of different LLMs to detect sexism in social media text using the EXIST 2024 tweet dataset. It includes annotations from six distinct profiles for each tweet, allowing us to evaluate to what extent LLMs can mimic these groups’ perceptions in sexism detection. Additionally, we analyze the demographic biases present in the models and conduct a statistical analysis to identify which demographic characteristics (age, gender) contribute most effectively to this task.Our results show that, while LLMs can to some extent detect sexism when considering the overall opinion of populations, they do not accurately replicate the diversity of perceptions among different demographic groups. This highlights the need for better-calibrated models that account for the diversity of perspectives across different populations.

AbstractTimely identification and analysis of disease transmission networks is becoming an emerging approach for understanding and mitigating the local spread of infectious diseases such as HIV and Hepatitis C. At the Centro de Investigación en Enfermedades Infecciosas, Mexico, we collect blood samples from local clinics, of people infected and sequence specific viral genes. Using bioinformatic tools (e.g. HIV-TRACE), we construct putative transmission networks by linking sequences with genetic distances below 1.5%. These networks allow us to identify clusters of related cases, which serve as epidemiological units of study. By analyzing cluster properties, such as size and connectivity, we investigate associated genetic, clinical, and demographic risk factors that contribute to the dynamics of viral spread.Despite the utility of these networks, the computational demands of network-building algorithms, particularly when applied to large datasets, remain a significant challenge. Artificial intelligence (AI) could offer transformative potential in this domain by improving the scalability, speed, and accuracy of network construction. Machine learning models could assist in optimizing sequence similarity thresholds, predicting cluster growth, and identifying key drivers of transmission.

The province of Entre Ríos (Argentina) lacks an automated model for electricity demand forecasting, which would notably enhance its management and daily operations. Accurate prediction remains a challenge due to the non-linear nature of electricity consumption. This work addresses this challenge using deep learning techniques with demonstrated success in time series prediction and growing application in energy forecastingWe propose an LSTM-based model for short-term electricity demand forecasting using a recurrent neural network (LSTM) capable of capturing temporal dependencies without manual feature extraction. The model takes as input a combination of meteorological variables (e.g., temperature, humidity) and energy-related variables (e.g., historical demand, time of day) to predict electricity demand. We evaluated various temporal window configurations and other hyperparameters to assess their impact on predictive performance. Results demonstrate that the LSTM model effectively captures temporal dependencies, providing accurate demand forecasts compared to classical methods. This study highlights the potential of deep learning for optimizing energy system operations and supporting decision-making in power management

Rumor detection has been widely studied from various perspectives and across multiple languages. However, for low-resource languages, such as Spanish, research remains limited due to the lack of labeled datasets necessary for training neural networks. In this work, we address this challenge by leveraging datasets from related tasks, including stance classification with English data and bot detection with both English and Spanish data. Our study explores how these datasets can support and transfer knowledge to improve rumor detection in Spanish.

Mangrove ecosystems are vital for biodiversity conservation and coastal protection but face severe threats from human activities, particularly in El Oro, Ecuador. This study presents MangroveNet, an advanced deep learning approach for mapping and monitoring mangrove forests using U2-Net and multispectral satellite data from Landsat and Sentinel-2. By integrating six spectral bands (RGB, NIR, SWIR-1, SWIR-2) and four multispectral indices (NDVI, NDWI, NDMI, MNDWI), our model achieves precise segmentation of mangrove areas. Unlike traditional methods, U2-Net's multi-scale feature extraction enhances accuracy in complex landscapes, enabling better tracking of deforestation and ecosystem changes. Performance is evaluated using Precision, Recall, Intersection over Union (IoU), and Dice score, ensuring robust segmentation. The results provide critical insights for conservation strategies, supporting reforestation efforts and sustainable management policies. This study demonstrates the potential of deep learning and remote sensing in environmental monitoring, offering a scalable solution for real-time mangrove ecosystem assessment.

Natural Semantic Metalanguage, theory formalized by Anna Wierzbicka, proposes a set of words whose meaning is universal across languages and cultures. Through this framework several datasets and academic papers have been published, along them the Longman Common Dictionary of English rewritten to be NSM-compliant. However, there is a gap in the literature regarding the complexity behind the induced network of semantic relationships. Through the study of this network, we may be able to create new architectures that are more efficient and allow us to inspire and analogous Physics Informed Neural Operator for Language Models.

The adoption of Open Science practices in Latin America presents both challenges and opportunities. This study examines the perceptions of 285 Spanish-speaking researchers and technicians regarding the barriers to Open Science implementation and its impact on the relationship between the scientific community and the public. To explore these perspectives, we apply a text mining approach using tools from the tidy ecosystem in R.Data were collected from registration forms for an Open Science training conducted in 2024 by MetaDocencia, based on NASA materials contextualized into Spanish. We employed classical NLP techniques, including frequency analysis, tf-idf computation, sentiment analysis, and topic modeling. The findings offer a structured overview of key challenges, such as institutional barriers and a lack of incentives, as well as expected benefits, including increased transparency and public trust. This study enhances our understanding of how the research community perceives the transition to Open Science and provides valuable insights for designing strategies that support its implementation.

We analyzed the performance of Convolutional Autoencoders in generating reduced-order representations the temperature field of 2D Rayleigh-Bénard flows at $\Pr=1$ and Rayleigh numbers extending from $10^6$ to $10^8$, capturing the range where the flow transitions to turbulence.We present a way of estimating the minimum number of dimensions needed by the Autoencoders to capture all the relevant physical scales of the data that is more apt for highly multiscale flows than previous criteria applied to lower dimensional systems.We compare our architecture with two regularized variants as well as with linear methods, and find that manually fixing the dimension of the latent space produces the best results.We show how the estimated minimum dimension presents a sharp increase around $Ra\sim 10^7$, when the flow starts to transition to turbulence. Furthermore, we show how this dimension does not follow the same scaling as the physically relevant scales, such as the dissipation lengthscale and the thermal boundary layer.

Principal Component Analysis (PCA) is a widely used technique for data processing, compression, and visualization. However, its reliance on linear projections and the absence of mechanisms to ensure fairness introduce significant limitations in many practical applications. To address these challenges, we propose a novel framework based on a mixture of Probabilistic Principal Component Analyzers (PPCA). This approach allows for the modeling of complex, nonlinear data structures while incorporating explicit controls for fairness across distinct groups. Our method leverages a probabilistic foundation to ensure robustness and interpretability, with applications spanning diverse domains where ethical and unbiased data representation is critical.

Es el resumen del proyecto de Google Assessing AI technology for biased and toxic behaviors is a crucial issue in developing human-centered applications based on AI. Existing approaches often fail to take into account culturally situated evaluations and lack representation from marginalized regions like Latin America. There are significant challenges in obtaining rich socio-cultural knowledge that requires community engagement. Empowering high-school teachers and their students to carry out such assessments themselves contributes to creating a responsible ecosystem, where they can understand the risks and make informed decisions on how to use AI. We are carrying out a large professional development (PD) course in Argentina with 30 university professors, 400 high school teachers, and 6K students that empower high school teachers and students to carry out such assessments systematically, as we teach them how they can use generative AI tools responsibly in their daily tasks. Teachers and students will generate outputs based on diverse prompts, analyze potential stereotypes and biases in these outputs, and validate findings with socio-cultural perspectives specific to Latin America. As a by-product of the exploration of biases carried out by teachers and students, datasets that represent harmful stereotypes are also produced, validated, and consolidated to assess biases and other risks in language technologies. This project proposes to scale this initiative to other 4 countries in Latin America: Brazil, México, Chile and Uruguay. Almost 400 million people, more than 60% of the latinamerican population, live in one of these 5 countries. We will investigate differences in methodology required by different cultures and contexts of application, and diversify and consolidate materials and methodologies adequate to those contexts. We will release the created datasets so that future work on society-centered AI can evaluate alignment in a more culturally aware way in Latin America.

The study of planetary transits is essential for detecting and characterizing exoplanets, enabling the analysis of their physical, atmospheric, and orbital properties, and improving our understanding of planetary system formation. As astronomical datasets, such as those from the operational Transiting Exoplanet Survey Satellite (TESS), survey a large fraction of the sky, they generate a vast database of photometric time series with the goal of identifying exoplanets. This work presents a Transformer-inspired neural network to detect exoplanet transit signals directly from Full Frame Image (FFI) light curves in the 30-minute cadence data, without relying on periodicity, prior transit parameters, or phase folding. The network is trained to learn characteristics of the transit signal, such as the dip shape, which helps distinguish planetary transits from other variability sources, such as eclipsing binaries. Our method successfully identified new planetary system candidates from TESS sectors 1-26 with a radius greater than 0.27 Jupiter radius. With inference times of just milliseconds per light curve, the TESS 30-minute cadence dataset can be processed efficiently, enabling real-time analysis of new data.