Automatically describing images using natural sentences is essential to help include people with vision impairment or low vision on the Internet, making it more inclusive and democratic. This task is known as image captioning and it is still an open challenge that requires understanding the semantic relations among objects in the image and transforming them into descriptions. Significant progress has been made in image captioning in recent years thanks to the availability of a large amount of labeled data in relevant datasets. However, most datasets only comprehend images annotated with English descriptions, whereas datasets with captions in other languages are scarce. In particular, Portuguese is a low-resource language; hence it has few publicly available datasets. Thus, to contribute to the community of Portuguese speakers, we have proposed the first large dataset for image captioning with descriptions in Portuguese, #PraCegoVer. Our dataset relies on public data published by followers of the PraCegoVer movement. Then, it comprehends 533,523 images with captions in Portuguese created explicitly for the audience with vision disabilities.

Airborne diseases are easy to spread in any population. The advent of COVID-19 showed us that we are not prepared to control them. The pandemic has drastically posed challenges to the daily functioning of public and private establishments. In general, while there have been several approaches to reduce the potential risk of spreading the virus, many of them rely on the commitment that people make, which - unfortunately - cannot be constant, for example, wearing a facemask in closed environment at all times or social distancing. In this work, we propose a stereo vision system to determine the risk of airborne disease spread in closed environments. We modify and implement the Wells-Riley epidemiological equation. The generated data from several devices is visible in a web platform to monitor multiple areas and locations. Finally, an OAK-D camera and a Jetson device are embedded in a end device meant to monitor a closed environment and send spread risk data continually to the web platform.

Affective computing has emerged as a new discipline that seeks to incorporate emotions into the realm of artificial intelligence. It is a growing field of research that combines computer science and engineering methods to automatically recognize and interpret human emotions. The use of physiological signals, such as electrodermal activity, has been a focus of recent research. Most of this work focuses on how to improve signal extraction or which machine learning models are best at recognizing emotions from this signal. However, the emotional models underlying such systems have received little attention in the literature. In response, the authors conducted a systematic review of the existing literature on automatic emotion recognition systems using electrodermal activity. The review considered both the machine learning models used and the affective and physiological components of each model. The PRISMA protocol was followed and the review was pre-registered in OSF. The results showed an increase in the use of dimensional models of emotion over time, but a corresponding lack of regression models. The authors conclude that it is essential to consider both affective and computational perspectives in the study of emotion in order to achieve a more comprehensive understanding of affective states.

"Multilingual generative language models are increasingly fluent in a large variety of languages. They are trained on huge corpora of multiple languages that enable powerful transfer from high-resource languages to low-resource ones. However, it is still unknown the sort of cultural biases that are induced in the predictions of these models. In this work, we analize the formality level of two multilingual language models: XGLM (Lin et al., 2021) and BLOOM (Scao et al., 2022), across five languages, namely Arabic, Bengali, English, French, and Spanish."

The timing and intensity of calling activity in anuran amphibians, which has a central role in sexual selection and reproduction, are largely controlled by climatic conditions such as environmental temperature and humidity. Therefore, climate change is predicted to induce shifts in calling behavior and breeding phenology, species traits that can be tracked using passive acoustic monitoring (PAM). To construct robust algorithms that allow classifying species calls in a long-term monitoring program, it is fundamental to design adequate datasets and benchmarks in the wild. We present a large-scale multi-species dataset of acoustics recordings of amphibians anuran from PAM recordings. The dataset comprises 27 hours of herpetologist annotations of 42 different species in different regions of Brazil. The classification task is unique and challenging due to the high species diversity, the long-tailed distribution, and frequent overlapping calls. We present a characterization of the challenges and a baseline model for the goals of the monitoring program. The dataset, including raw recordings, preprocessing code, and baseline code, is made available to promote collaboration between machine learning researchers and ecologists in solving the classification challenges toward understanding the effects of global change on biodiversity.

"Automated Machine Learning (AutoML) has become increasingly popular in recent years due to its ability to reduce the amount of time and expertise required to design and develop machine learning systems. This is very important for the practice of machine learning, as it allows building strong baselines quickly, improving the efficiency of data scientists, and reducing the time to production. However, despite the advantages of AutoML, it faces several challenges when it comes to real-world applications. In AutoMeLi, our internal AutoML tool, there are a number of predefined steps consisting of data preprocessing, feature engineering, and an estimator. Whenever a new step is added, the number of possible pipelines grows exponentially. In order to overcome this situation, we developed GramML, our solution for defining and exploring a grammar-based search space. This is a model-free reinforcement learning approach based on an adaptation of the Monte Carlo tree search (MCTS) algorithm and context-free grammars, which has shown state-of-the-art performance on the OpenML-CC18 benchmark."

Electronic support measures involve the collection and characterization of radiated electromagnetic energy sources for situational awareness and decision-making, and thus they are an essential element of electronic warfare capabilities. In highly congested and contested environments, it is fundamental to characterize pulse-Doppler radar emitters for friend-or-foe classification. Deep learning applications for radar signal processing are promising for their potential to process in parallel large amounts of data coming from complex scenarios. Unfortunately the development of discriminative models for pulse-Doppler radar data is hindered by the scarcity of relevant recordings, as military emitters are aware of the electronic intelligence activities of other agents and limit their emissions. We use Variational Autoencoders (VAE) for learning latent variable models of the emitters-of-interest, leveraging them as data augmentation tools of these emitters' pulse descriptor words to power our deep learning radar processing applications. VAEs crystallize the information provided by newly gathered data and become a powerful tool for machine-learning-based, automated electronic intelligence.

This work proposes an automatic segmentation approach to detect White Matter Hyperintensities (WMH) using Fluid Attenuated Inversion Recovery (FLAIR) and the corresponding T1 weighted MRI images. In this work, we used the Swin U-Net architecture based on Transformers and compared its performance with two currently reported CNN U-Nets architectures. Sixty pairs of images and their corresponding ground truth labels were used from the MICCAI challenge. The following metrics were obtained: Dice similarity score of 0.80; lesion F1 score of 0.63; Hausdorff distance of 3.16 mm; and, 16.93 average volume difference, locating the Swin U-Net second in the ranking of the tested algorithms. However, the computational resources needed to process the imaging data were the lowest compared to the previous U-Nets. Therefore, the Swin U-Net architecture shows potential to become a promising and fast tool for segmenting medical images.

"Classically, it has been considered that the only groups of birds that exhibit vocal learning are oscines (""traditional"" songbirds, of the Passeri suborder and order Passeriformes), parrots (order Psittaciformes) and hummingbirds (order Apodiformes, family Trochilidae); who are postulated to have evolved these traits independently. On the other hand, the sister suborder of oscines, predominantly present in South America, the suboscines (order Passeriformes, suborder Tyranni), presents birds considered to be innate vocalizers without plasticity, with vocal organs with little musculature and innervation compared to the previous groups. However, recent studies in suboscines have shown cases of birds with variability in the use and form of their vocalizations, and even in specific cases, presenting plasticity and vocal learning. While there is still very little information on the vocal ecology of suboscines, in this work we investigated the vocal behavior of three Chilean suboscines using directional field recordings, to reveal their vocal repertoires. Using unsupervised machine learning techniques, recordings of the suboscines species Turca (Pteroptochos megapodios), and Fiofio (Elaenia chilensis), were analyzed in search of characteristic vocal patterns to distinguish the main vocalizations and the qualities that allow them to be classified."

Automatic cancer diagnosis based on RNA-Seq profiles is at the intersection of transcriptome analysis and machine learning. Methods developed for this task could be a valuable support in clinical practice and provide insights into the cancer causal mechanisms. To correctly approach this problem, the largest existing resource (The Cancer Genome Atlas) must be complemented with healthy tissue samples from the Genotype-Tissue Expression project. In this work, we empirically prove that previous approaches to joining these databases suffer from translation biases and correct them using batch z-score normalization. Moreover, we propose CanDLE, a multinomial logistic regression model that achieves state of the art performance in multilabel cancer/healthy tissue type classification (94.1% balanced accuracy) and all-vs-one cancer type detection (78.0% average max F1).

"We present Class Based-Active Queue Management (CB-AQM), a novel approach for QoS in MANETs, addressing the problem of guaranteeing a minimum operational infrastructure for critical communications over Disas- ter Area Networks while allowing non-critical clients (i.e. people living in the disaster area) to use the communications infrastructure under a best-effort scheme. Non-prioritized traffic is expected to absorb the drawbacks of poor connectivity during eventual periods of traffic overload."

In this paper we study the training of sparse neural networks and show that constrained formulations can deliver improved tunability and hyperparameter interpretability compared to the commonly used L0-penalty regularization. Our proposal reliably achieves arbitrary sparsity targets while retaining high accuracy, and scales successfully to large residual models -- and all this with just a negligible computational overhead!

"Cooper (https://github.com/cooper-org/cooper) is a general-purpose, deep learning-first constrained optimization library in Pytorch. Cooper is (almost!) seamlessly integrated with Pytorch and preserves the usual loss ➡️ backward ➡️ step workflow. If you are already familiar with Pytorch, using Cooper will be a breeze! This library aims to encourage and facilitate the study of constrained optimization problems in deep learning. Cooper focuses on non-convex constrained optimization problems for which the loss or constraints are not necessarily “nicely behaved” or “theoretically tractable”. Moreover, Cooper has been designed to play nicely with mini-batched/stochastic estimates for the objective and constraint functions. Cooper implements several popular constrained optimization protocols so you can focus on your project, while we handle the nitty-gritty behind the scenes. "

Social media data has emerged as a useful source of timely information about real-world crisis events. Several studies have addressed the automatic detection of crisis-related messages to contribute to disaster management and humanitarian assistance. However, most of them have focused on a particular language (usually English) or domain (type of event), which limits their applicability to other contexts. In this work, we study the task of automatically classifying messages that are related to crisis events by leveraging cross-language and cross-domain labeled data. Our goal is to make use of labeled data from high-resource languages to classify messages from other (low-resource) languages and/or of new (previously unseen) types of crises. We proposed an experimental framework based on combinations of multilingual data representations and knowledge transfer scenarios. Our experimental results showed that it is possible to leverage English data to classify the same domain in other languages, such as Spanish and Italian (80.0% F1-score). Also, to classify messages from new domains (80.0% F1-score) in a multilingual scenario. Overall, our work contributes to mitigating cold-start situations in emergency events, when time is of essence.

In Argentina, Electronic Health Records (EHR) have been continuously implemented increasing the amount of this type of data. Ethical considerations arise for their reuse to address secondary research, public health, management, and policy-making questions. Health data are sensitive data according to national and international regulations (HIPAA, GDPR, etc.) because they can significantly impact people’s lives. Thus, having tools for effectively eliminating protected personal information (PPI) that could allow patient identification is a must. But anonymization of free text in EHR is a challenging problem because it is full of peculiarities (words outside common vocabulary, ambiguity, etc.). We are presenting our experience in developing a de-identification algorithm for free text in the EHR of a province in Argentina. We found that it is not clear to humans what information is PPI during a manual anonymization task. As expected, automatic processes also miss cases of PPI, even more than humans do. However, a simple, rule-based approach can do a good job in removing most of the PPI, outperforming a more sophisticated, machine-learning approach for low-resources contexts. Although no process can guarantee anonymization, our method can mitigate the impact of possible data breaches from highly sensitive information.

"The development of 5G networks has opened new opportunities for the integration of a wide range of applications, from the Internet of Things (IoT) to autonomous vehicles, requiring high-speed and reliable communication. However, these new use cases also impose new challenges for network operators, particularly in terms of scheduling, which is critical to ensuring optimal performance and quality of service. Traditional scheduling algorithms based on heuristics have limitations when it comes to addressing the complexity and dynamicity of 5G networks, particularly in terms of the massive number of devices and the heterogeneity of their requirements. To address these challenges, researchers have been exploring the potential of reinforcement learning (RL) as a promising approach for scheduling in 5G networks. RL is a subfield of machine learning that involves training agents to learn optimal actions through trial-and-error interactions with their environment. By leveraging RL, scheduling in 5G networks can be optimized in a data-driven manner that adapts to the changing network conditions and user demands. In this poster, we not only present a review of modern scheduling algorithms for 5G networks, but also demonstrate their performance through implementation in a 5G simulator. "

Edge-adapted methods have been introduced in the context of image processing to reconstruct high-resolution images from coarser cell averages. In particular, when images consist of piecewise smooth functions, the interfaces can be approximated by a pre-specified functional class (lines, circle arcs, etc) through optimization (LVIRA) or specific preprocessing (ENO-EA). In this talk, we will first explore some theoretical aspects of these nonlinear approximation spaces that are useful in the context of inverse problems. Secondly we will show an extension of the ENO-EA approach to polynomials of degree higher than 1 and compare this algebraic approach to that introduced in (LVIRA) as well as to learning-based methods [B. Desprs, H. Jourdren 2020] in which an artificial neural network (NN) (or in principle any other non linear sufficiently rich function family) is used to attain the same goal.

"Deep Learning has shown outstanding results in computer vision tasks, and healthcare is no exception. Deep Learning (DL) can assist dermatologists in early skin cancer diagnosis, saving many lives. However, there is no straightforward way to map out the decision-making process of DL models. For skin cancer predictions, it is not enough to have good accuracy. Understanding the model's behavior is needed to implement it clinically and get reliable predictions. We propose desiderata for explanations in skin-lesion models and present a study about how eXplainable Artificial Intelligence (XAI) is currently used for skin lesions. We analyzed seven methods (four based on pixel-attribution and three high-level concepts): Grad-CAM, Score-CAM, LIME, SHAP, ACE, ICE, CME for two deep neural networks, Inception-v4 and ResNet-50, trained on the International Skin Imaging Collaboration Archive (ISIC). Our findings indicate that while these techniques effectively show what the model is looking at to make its prediction, the obtained explanations are not complete enough to get transparency into the skin-lesion models."

" Chain-of-thought (CoT) prompting has been purported to improve the performance of large language models on a variety of tasks. While initially being limited to massive language models (60B+ parameters), recent work has shown that this capability can be unlocked in smaller models via strategies such as instruction fine-tuning including explanations for ground-truth answers. However, it is not immediately clear whether chain-of-thought prompting has a positive impact on the performance of smaller models on tasks such as commonsense reasoning. As part of this work, we evaluate six medium-sized language models on five commonsense question answering datasets. Our results suggest that chain-of-thought prompting often does not improve the performance of such models when compared to direct answering. "

The social effects of the development and deployment of artificial intelligence require multidisciplinary approaches to analyze AI as a socio-technical artifact embedded in social relations of power. Our effort aims to propose the need to promote multidisciplinary research from a decolonial, feminist, and anti-capitalist perspective in Latin America and the Caribbean to provide critical views on technologies in general and intelligent systems in particular, necessary to promote technological development alternatives that respond to the realities and needs of communities in the region, considering the inequality gaps, cultural, linguistic, gender, racial, social and educational differences in political contexts of corruption and impunity. Based on the experience of the Feminist Artificial Intelligence Research Network in Latin America and the Caribbean, we present a research agenda for the region, concerns, and preliminary reflections.

The online sharing and viewing of Child Sexual Abuse Material (CSAM) are growing fast, such that human experts can no longer handle manual inspection. However, the automatic classification of CSAM is a challenging field of research, due to the inaccessibility of target data that is — and should forever be — private and in sole possession of law enforcement agencies. In this work, we inspect a broad set of visual features from the literature on CSAM detection and reports from law enforcement institutions. First we provide a structured overview of content level information from a real database, to aid researchers in drawing insights from unseen data and safely providing further understanding of CSAM images. Later in this project we intend to assess the the predictive power of the inspected dimensions, by producing several machine learning models from a diverse combination of this wide set of features. We demonstrate our proposal on the Region-based annotated Child Pornography Dataset (RCPD), one of the few CSAM benchmarks in the literature, produced in partnership with Brazil's Federal Police. Although limited in several senses, we argue that automatic signals can highlight important aspects of such datasets, which is valuable when data can not be disclosed.

Kernel methods have been widely used in pattern recognition for supervised learning tasks such as sufficient dimension reduction, supervised feature selection and graph classification. These methods methods are useful in statistical learning applications affected by the curse of dimensionality such as biomedical data. In this work I introduce a novel approach to incoporate distribution structure as training labels in order to improve the supervised task.

Smoothness and low dimensional structures play central roles in improving generalization and stability in learning and statistics. The combination of these properties has led to many advances in semi-supervised learning, generative modeling, and control of dynamical systems. However, learning smooth functions is generally challenging, except in simple cases such as learning linear or kernel models. Typical methods are either too conservative, relying on crude upper bounds such as spectral normalization, too lax, penalizing smoothness on average, or too computationally intensive, requiring the solution of large-scale semi-definite programs. These issues are only exacerbated when trying to simultaneously exploit low dimensionality using, e.g., manifolds. This work proposes to overcome these obstacles by combining techniques from semi-infinite constrained learning and manifold regularization. To do so, it shows that, under typical conditions, the problem of learning a Lipschitz continuous function on a manifold is equivalent to a dynamically weighted manifold regularization problem. This observation leads to a practical algorithm based on a weighted Laplacian penalty whose weights are adapted using stochastic gradient techniques. We prove that, under mild conditions, this method estimates the Lipschitz constant of the solution, learning a globally smooth solution as a byproduct.

SiO2 micro- and nanoparticles have diverse applications. A calibrated, monodisperse size is a requirement for the use of these particles in nanomedicine or for the design of new composite materials. Stöber synthesis provides such particles, is robust and scalable. Its main disadvantage is to reach a size that has not been synthesized before due to the complexity of the synthesis process. Machine learning methods are beginning to be used to assist in the synthesis of materials and can be a very useful tool to predict the outcome of Stöber synthesis. These methods require data, ideally abundant and of good quality, in order to learn and make good predictions. The aim of this work is to explore the limitations of working with secondary data reported in the literature.

"Executive functions (EF) are a class of processes critical for purposeful, goal-directed behavior. Cognitive training has been studied and applied for more than 25 years in hugely diverse backgrounds. In spite of the accumulated evidence of its positive impact in cognition, there are still reports in the literature that claim that the potential benefits of training are not generalizable. Recently, research regards individual differences as one of the possible causes of these inconsistencies. We consider it is time to start using individual differences as information to guide us towards finding better strategies for each person instead of inevitable experimental noise. In a previous study, we trained a classifier algorithm that successfully predicted whether a subject would benefit, or not, from a fixed training approach, based on his/her performance in previous cognitive tests (accuracy= .67, AUC=.707). In this study we dissect our previous result by training an algorithm to predict specific improvements in Working Memory abilities, which resulted in even better performance (accuracy=.79). Results indicate that in order to improve working memory children need prior solid attentional resources. We believe that understanding the implications of these results is the key to understanding how to personalize cognitive training."

Typically, in data augmentation for deep learning (DL) new training samples are generated by suitable modifications of the original data, and then the model is trained via mini-batch stochastic gradient descent. In the context of DL for brain-computer interfaces (BCIs) using simulated electroencephalography (EEG) can be a better approach for data augmentation. However, random sampling might not be appropriate for BCIs applications due to data scarcity. Moreover, synthesized data can be thought of as samples from a different domain than real data and this fact must also be taken into account when sampling for the construction of mini-batches. Here, data augmentation samples consist of subject-specific simulated MI-EEG trials, and different mini-batch sample selection techniques are evaluated: 1) random sampling, 2) domain-aware sampling (DAS), in which all mini-batches have the same ratio of real/augmented samples, 3) stratified sampling (SS), in which mini-batches are stratified by output class, 4) stratified domain-aware sampling (SDAS), which combines DAS and SS. Results demonstrated that data augmentation strategies with mini-batch sample selection techniques that preserve class balance and the ratio of real/augmented data (SDAS) improve DL-based MI decoding, significantly outperforming baseline performance.

" The development of in vivo registration technologies has pushed systems neuroscience field growth by allowing the simultaneous recording of multiple data modalities in complex experimental setups. These modalities are a set of time series signals from brain activity (from silicon probes, neuropixels, or calcium imaging) in combination with behavioral readouts (facial expression or navigation videos) and usually categorical external stimuli (external sounds, a set of somatosensory stimuli). Here we will present two examples, for both single or combined data modalities analysis, where standard machine learning tools are used to extract relevant features that help to describe brain activity. At the same time, we will discuss why and how the development of specific models for these datasets could help to better understand the underlying process generating complex behavior and complex brain signals. The first example consists of the classification of electrophysiological recordings into different oscillatory types (one data modality). The second example is related to the prediction of behavioral/stimuli based on calcium imaging signals (a combination of data modalities). "

Enabling low precision implementations of deep learning models, without considerable performance degradation, is necessary in resource and latency constrained settings. Moreover, exploiting the differences in sensitivity to quantization across layers can allow mixed precision implementations to achieve a considerably better computation performance trade-off. However, backpropagating through the quantization operation requires introducing gradient approximations, and choosing which layers to quantize is challenging for modern architectures due to the large search space. In this work, we present a constrained learning approach to quantization aware training. We formulate low precision supervised learning as a constrained optimization problem, and show that despite its non-convexity, the resulting problem is strongly dual and does away with gradient estimations. Furthermore, we show that dual variables indicate the sensitivity of the objective with respect to constraint perturbations. We demonstrate that the proposed approach exhibits competitive performance in image classification tasks, and leverage the sensitivity result to apply layer selective quantization based on the value of dual variables, leading to considerable performance improvements.

Non-Intrusive Load Monitoring (NILM) is a technique for disaggregating the energy consumption of individual appliances from a single aggregate signal. This study presents a novel approach for NILM at very low frequencies, focusing on electric water heaters and electric vehicles in Uruguay. The study uses a dataset provided by the Uruguayan electricity company. The data is divided into daily segments, allowing a Neural Network to learn the patterns of daily energy consumption for each appliance. Our approach is one of the first to work with electric water heaters and electric vehicles at this frequency. In addition, the study also involves a classification task for identifying different appliances. The method achieves a mean absolute error of 100 W, which is a reasonable performance at very low frequencies. The proposed approach can be applied to a wide range of NILM applications in Uruguay and similar countries.

Perception of toxicity evolves over time and often differs between geographies and cultural backgrounds. Similarly, black-box APIs for detecting toxicity, such as the PerspectiveAPI, are not static but frequently retrained to change and improve on their unattended weaknesses and biases. Although prone to errors, these tools are widely used in safety research to benchmark reductions in harm for large-scale evaluations and baseline definitions. Their evolving nature, however, poses a challenge given their widespread adoption. The RealToxicityPrompts (RTP) dataset, developed to study language models' degeneration into toxicity, is composed of prompts, continuations, and the toxicity score for each, obtained from the PerspectiveAPI. Our work shows that toxicity score distributions for prompts and continuations have changed considerably from when the dataset was released to this day. This implies that previous work that either inherited automatic toxicity results from baseline papers or used the original RTP scores to stratify freshly scored sequences, might not have been accurate as per the API's definition of what toxicity was at the moment. Our findings suggest caution in applying apple-to-apple comparisons between works and call for more structured approaches to evaluating toxicity over time.

Armed robbery and some other weapon related crimes are one of the most important social problems in Central América. Using a custom manually collected dataset from news and social media a computer vision algorithm was created using deep learning to identify potential robberies, crimes, guns and dangerous situations, the model is optimized for Google Coral boards to run on-device/edge inference on real time, this can be adapted to house cameras, car dashcams and potentially send alerts to nearby users.

"The Semantic Web has the potential for serious impact on how personal data is collected and used. Automated and autonomous data collection has grown exponentially over the past few years. More personal and highly detailed data can be collected and analyzed than at any other time in history. Differential privacy offers strong and robust privacy guarantees. It is based on randomised aggregatable response in answers which enables good estimations of accurate population statistics while preserving the privacy of the individual respondents. Our purpose is to prevent the risk of associating sensitive information with the identities of data owners, and, to measure how effective differential privacy techniques are at protecting privacy within larger ontologies. For this purpose, in this Ph.D. project, we aim to propose a differential privacy model for ontologies. Our main contributions are: (1) development of a privacy-preserving model via aggregation synthetic data that satisfy differential privacy; (2) analysis of the feasibility, reliability, and performance of our proposal based on case studies;"

With the advance of multimodal training strategies, some of them have become very popular, such as text-guided training of image models. One of the most influential works in this field is the CLIP (Contrastive Language-Image Pre-Training) which proposes a pre-training to pair captions with their most suitable image. This model is trained with images and respective captions extracted from the internet, avoiding the need for human supervision and annotation of specific concepts represented in the images. This pre-training mode gives great results even in zero-shot scenarios. In this work, we built a large dataset entirely in Portuguese, by merging multiple sets originally in Portuguese and various translated datasets widely used in English. From this set, we train and evaluate a CLIP structure entirely in Portuguese, evaluating aspects such as different encoders, the impact of translations, and duplicate captions. The results are compared with the original CLIP, the multilingual CLIP, and specific language versions such as Chinese CLIP and Italian CLIP.

"Protein Electron Microscopy (EM) maps are critical in determining the three-dimensional structures of bio-molecules, including proteins and their interactions. The task of identifying regions that correspond to specific proteins is challenging, but crucial for gaining insight into their function and designing drugs to enhance or suppress their processes. Conventional methods of protein EM map segmentation use segmentation algorithms that assign a voxel to a region, but can result in difficulties in obtaining a segmentation that maps each region to a single protein unit. Our approach extends the region optimization technique by incorporating an interactive mechanism that allows for user guidance of the segmentation process. Our deep learning model is trained on a dataset of protein EM maps and uses a convolutional neural network architecture. Results show that our approach has potential to provide efficient and accurate segmentation compared to traditional methods and is comparable to state-of-the-art approaches."

"On their surface, proteins are shaped into numerous cavities and protrusions that create unique microenvironments for ligand binding, catalysis, or other biologically relevant processes. Despite their biological and biotechnological relevance and their potential impact on various research areas in medicine, drug design, and evolutionary biology, there is still no exhaustive and comprehensive research on protein cavities combining sequential, structural, and evolutionary perspectives. We propose a representative set of features that enable machine learning, such as clustering and classifying the different types of protein cavities into functional categories. To this end, we combined classical bioinformatics techniques for molecular characterization with unsupervised learning techniques. We created and curated a dataset containing all proteins with known structures and their features, which is currently available in CaviDB. In this dataset, we computationally predicted the cavities present in each of the proteins and characterized them structurally and sequentially using a set of 47 features. We explored the data using nonlinear projections with UMAP and Self-Organized Maps. Using biological information such as volume, drug content, and polarity, we were able to cluster representative cavities and use them to reconstruct protein families."

"Missing data is abundant in predictive tasks. Typical approaches assume that the missingness process is ignorable or non-informative and handle missing data either by marginalization or heuristically. Yet, data is often missing in a non-ignorable way, which introduces bias in prediction if not treated properly. In this work, we develop a new method to perform tractable predictive inference under non-ignorable missing data using probabilistic circuits derived from Decision Tree Classifiers and a partially specified response model of missingness. We show empirically that our method delivers less biased (probabilistic) classifications than approaches that assume missing at random and are more determinate than similar existing overcautious approaches."

"An essential resource in the preservation of earth's biodiversity is keeping large natural areas protected. Unfortunately, sites of ecological interest are constantly threatened by illegal actors and the manpower allocated to monitor and safeguard these spaces is often insufficient. People in charge of taking care of those areas have to optimally allocate the patrolling resources in extensive tracts of land and are often in great disadvantage against the attackers. This problem has been previously studied in the literature and most of the work focuses on open spaces like the African Savannah or requires a discretization of the protected area. These approaches do not capture the reality of South American parks where, due to the density of the vegetation and the ruggedness of the terrain, traveling is done only over a limited collection of available trails (a graph). The problem addressed in this work is the design of near-optimal patrol schedules for rangers in such graphs. We illustrate our results in the trail map of Jamacoaque, a natural reserve in Ecuador. "

The conventional ConvNets are easily fooled even by human-imperceptible corruptions. In this poster, we will present a new trainable entry layer named the monogenic layer which in concurrence with different ConvNets outperforms regular ConvNets and adversarial training in front of large illumination changes.

Accurate diagnosis of thyroid nodules is crucial in clinical practice as it can help in early detection of thyroid cancer. In this study, a machine learning approach for the diagnosis of thyroid nodules in ultrasound images is proposed. The proposed method uses two modules, a Convolutional Neural Network (CNN) and a Decision Tree (ML), to perform the classification task. The transfer learning technique is applied to overcome the challenge of small data sets. The used dataset consists of 124 patients, with 18 having thyroid nodules and 106 without, and comparative experiments were performed to evaluate the performance of the proposed method. The results show that the proposed method has a significantly high thyroid nodule detection rate compared to other existing methods.

This work represents an expansion of the recently proposed benchmark framework for evaluating visual search models in natural scenes: ViSioNS. Our previous study provided a unified format and criteria for evaluating the efficiency and similarity of different state-of-the-art visual search models to human behavior. In this expanded work, we introduce two additional models: EccNet and an adaptation of a more classical model, Entropy-Limit Minimization (ELM). By incorporating these models, we are able to provide a more comprehensive evaluation of the performance of different state-of-the-art visual search models. Our results highlight the limitations of current models and demonstrate the value of integrating different approaches with unified criteria to develop more accurate and efficient visual search algorithms. Furthermore, our work contributes to addressing the urgent need for benchmarking data and metrics to support the development of more general human visual search computational models.

Melanoma is the deadliest form of cancer. Multiplexed immunofluorescence is a technique to analyze several markers which can be used to analyze dendritic cells encarged of activating cytotoxic CD8+ T lymphocytes responsible for killing tumor cells. Analyzing multiplexed images can be time consuming, requiring a large amount of manual labor. Here, we evaluated several machine learning approaches to automatically analyze multiplexed images in a reproducible and accessible manner.

Plant diseases can cause significant yield and quality losses, and some diseases can even produce toxins that are hazardous to human health. Management measures against these diseases depend on adequate and fast classification. Deep learning, applied to computer vision, is a state-of-the-art method with many successes in agriculture and medicine where disease classification is a very important task. Olive leaf diseases are a problem that threatens the quality of the harvest in the producers, so the objective of this work is to classify olive leaf diseases with Deep Learning in the olive harvests of the La Yarada-Los Palos in the Tacna region, Peru. For this purpose, a novel and public RGB image dataset were elaborated with leaves of olive plants affected by the most common diseases: virosis, fumagina, and nutritional deficiencies. Furthermore, we conduct extensive comparative experimental studies using all possible configurations of Data Augmentation, Transfer Learning, and Fine-Tuning techniques to train a modified VGG16 architecture for olive leaf disease classification, achieving 100%, 98.67%, and 100% accuracy in the training, validation, and test sets respectively. Thus demonstrating experimentally good compatibility on Transfer Learning, very good ability of the pre-trained model to adapt the feature extraction to our dataset using Fine-Tuning, and very good support provided by the Data Augmentation technique to avoid model overfitting. Also was found experimentally that the use of Fine-Tuning without Data Augmentation is not effective. The reproducibility of the experiments is ensured by making public the novel dataset and final model on our GitHub.

We outline the research lines of the Knowledge Management and Information Retrieval Research Group from the Institute for Computer Science and Engineering (ICIC CONICET UNS). Our main projects include learning causal models from digital media, building stance and opinion trees from social media, and the design and application of topic-based search methods in different domains.

Laptops are used everyday for work, leisure and other activities. People with motor disabilities, such as paralysis, are not able to fully utilize them. Digital solutions have been developed to help on these situations. Eye tracking systems allow users to command their devices only by their eyes movements. This technology is expensive for general public and consists of specialized hardware and software. The main objective of this work is to understand how these systems work and try to develop a prototype which could allow more people to use a laptop in a more affordable way.

Multi-class learning (MCL) methods perform Automatic Text Classification (ATC), which requires labeling for all classes. MCL fails when there is no well-defined information about the classes and requires a great effort to label instances. One-Class Learning (OCL) can mitigate these limitations since the training only has instances from one class, reducing the labeling effort and making the ATC more appropriate for open-domain applications. However, OCL is more challenging due to the lack of counterexamples for model training, requiring more robust representations. However, most studies use unimodal representations, even though different domains contain other information that can be used as modalities. Thus, this study proposes the Multimodal Variational Autoencoder (MVAE). MVAE is a multimodal method that learns a new representation from more than one modality, capturing the characteristics of the interest class in an adequate way. MVAE explores semantic, density, linguistic, and spatial information modalities. We explore MVAE with a few labeled instances in fake news, relevant reviews, and interest event detections. MVAE outperforms other baselines commonly used in the one-class text learning literature with statical significance differences.

The advancement of the Natural Language Processing field has allowed for the development of causal language models with a high capacity for generating text. For several years, the field of Cognitive Neuroscience has been using these models to better understand cognitive processes. In previous works, we found that models such as Ngrams and those based on LSTM networks are capable of partially modeling the predictability of words (cloze-Pred), particularly when used as a co-variable to explain the eye movements of people reading natural texts. In this work, we deepen this line of research using models based on GPT2. The results show that this architecture achieves better modeling of cloze-Pred than its predecessors. We also found and that fine-tuning the model with a domain-specific corpus generates its predictions to be less based on lexical frequency. Here we also discuss how to generate an multi-domain corpus of text in "Español Rioplantese".

The increased sophistication of next generation mobile networks such as 5G and beyond or the Flying Ad-hoc NETworks, and the plethora of devices and novel use cases to be supported by these networks, make of the already complex problem of resource allocation in wireless networks a paramount challenge. We address the specific problem of user association, a largely explored yet open resource allocation problem in wireless systems. We introduce GROWS, a deep reinforcement learning (DRL) driven approach to efficiently assign mobile users to base stations, which combines a well-known extension of Deep Q Networks (DQNs) with Graph Neural Networks (GNNs) to better model the function of expected rewards. By leveraging the benefits of combining reinforcement learning and graph representation learning, GROWS is able to learn a user association policy which improves over currently applied assignation heuristics, both in throughput utility as in diminishing user rejection.

"Automated decision-making systems, specially those based on natural language processing, are pervasive in our lives. They are not only behind the internet search engines we use daily, but also take more critical roles: selecting candidates for a job, determining suspects of a crime, diagnosing autism and more. Such automated sys- tems make errors, which may be harmful in many ways, be it because of the severity of the consequences (as in health issues) or because of the sheer number of people they affect. When errors made by an automated system affect a population more than other we call the system biased. Most modern natural language technologies are based on artifacts obtained from enormous volumes of text using machine learning, namely language models and word embeddings. Since they are created applying subsymbolic machine learning, mostly artificial neural networks, they are opaque and practically uninterpretable by direct inspection, thus making it very difficult to audit them. In this poster we present a methodology that spells out how social scientists, domain experts, and machine learning experts can collaboratively explore biases and harm- ful stereotypes in word embeddings and large language models. "

There are several challenges in learning sequence representations, especially in the context of few labeled data, high class imbalances and domain variability. In bioinformatics, the determination of secondary structures from biological sequences (such as RNA) is a very costly process, which cannot be scaled up efficiently, limiting our ability to functionally characterize such molecules. Non-coding RNAs are relevant for numerous biological processes in medical and agroindustrial applications. Computational methods are promising for the prediction of RNA structures, but show limited capacity for modeling their wide structural diversity. We present new end-to-end approaches for secondary structure prediction from the sequence alone. To harness larger datasets with unlabeled sequences, a self-supervised encoder is proposed. Then, an architecture based on RenNet is used to encode information about each sequence position and its neighbors, and then use element pair information to predict a connection matrix. We have compared several recent methods for secondary structure prediction, with special focus on how to measure generalization capabilities, using benchmark datasets and experimentally validated sequences. By using biophysical constraints to guide the learning, results are improved for different types of RNAs.

"Investigaciones recientes sugieren que embeddings como word2vec pueden codificar información sobre la frecuencia de las palabras. En este trabjo decidimos estudiar más exhaustivamente este fenómeno poco estudiado. Aquí encontramos que Skip-gram con negative sampling (SGNS), GloVe y FastText tienden a producir una mayor similitud semántica entre las palabras de alta frecuencia (A). La asociación entre frecuencia y similitud también aparece cuando las palabras se permutan aleatoriamente antes de entrenar — esto demuestra que los patrones que encontramos son un artefacto producido por los embeddings, los cuales tienen la capacidad de codificar la frecuencia de las palabras, además de su semántica (B). Por último, hacemos un experimento para ilustrar que la frecuencia de las palabras puede influir mucho en la medición de sesgos de género con embeddings (C)."

This poster is divided in two parts. The first one is about Cohere’s latest summarization model, currently available in Public Beta. This model was able to overcome LLMs input token limitations and provided users with control over the generated summaries’ length, abstraction level, and output format. The model was able to effectively summarize longer texts, such as academic papers, and is applicable to various use cases, including news and articles. The second part is about Cohere’s Discord bot that allows users to provide feedback on the quality of texts generated by our latest models, and generate data that will be used to fine-tune LLMs. Users who engage will have early access to Cohere's latest converse models and will be an important part on improving them.

Aprendizaje por Refuerzo aplicado al despacho óptimo de energía del sistema interconectado nacional utilizando SimSEE.

Event level data powers all current machine learning models. However, as users become more privacy-conscious, availability to event level data becomes sparser. Instead, to protect data privacy, training of models can be done using aggregate data. In this poster we describe techniques for training models when labels can only be observed in aggregate. We demonstrate that with a simple trick, one can adapt pipelines that consume event level data to instead train using aggregate label information.

La Estrategia de IA surge como herramienta para la transformación digital que se desarrolla con el propósito de promover y fortalecer su uso responsable en la Administración Pública. Para ello se definen 9 principios generales, pilares, objetivos y líneas de acción que faciliten y fomenten el uso responsable de esta tecnología en el ámbito público.

Use the data from the telemetering as well as static data of the clients and the electric grid in order to improve and validate the topology of the electrical grid as well as determine the way of using the air conditioning in our clients.

" The field of Reinforcement Learning (RL) has gained significant attention in recent years due to its ability to optimize control using reward signals. In this study, we focus on a monitoring problem, which involves monitoring regions with occupation restrictions using one autonomous agent. The problem is modeled as an RL problem, where the dual variables act as a reward signal. To solve the problem, we propose a novel parameterization approach through a neural network that processes both primal and dual variables in parallel. By incorporating this structural innovation, the network learns to select navigation policies based on the degree of constraint satisfaction. Moreover, this satisfaction level is observed in real-time through the dual variables. To evaluate the proposed approach, we conduct experiments with a single agent monitoring multiple regions. The results indicate that the approach can efficiently optimize the control for monitoring the restricted regions with a high degree of constraint satisfaction. Our proposed approach has significant implications in monitoring applications, where the dual variables can be used as a measure of satisfaction and the optimization is based on reward signals in RL. Our results also highlight the potential of using RL techniques with dual variables for monitoring problems."