Dr. Luis Caro Saldivia

Technology played a crucial role in combating the COVID pandemic, both in the rapid development of vaccines and the early detection of the virus. Consequently, numerous studies in the medical field have focused on leveraging the power of artificial intelligence for COVID-19 detection. However, in the medical domain, it is essential to have a clear understanding of the processes and algorithms used in decision-making, as these directly impact people’s health. Therefore, efforts have been made to implement explainable artificial intelligence techniques, enabling humans to understand and explain the deep learning algorithms used in disease detection. In this work, we present an approach to detecting COVID-19 in chest X-rays that combines reconstruction-based anomaly discov- ery with perturbation-based attribution. Specifically, we use a Variational Autoencoder trained on healthy lungs to identify lung anomalies, and we additionally evaluate a Per- ceptual Autoencoder (PAE) that incorporates a perceptual loss to produce sharper recon- structions and more contrastive residual maps. These signals are then used to highlight critical image evidence through both patch-level scoring (grid-based occlusion) and pixel- level masking derived from reconstruction-error maps, enabling healthcare professionals to localize relevant regions more effectively. Moreover, the proposed framework provides clearer explanations of the model’s decisions by quantifying the prediction change after occluding the identified regions, reducing the complexity of the “black boxes” generated by deep learning neural networks. With this methodology, we aim to improve the effec- tiveness and reliability of early COVID-19 detection through chest X-rays.

Demand for image inspection in dermatology is growing, leading to the exploration of deep learning techniques and algorithms to enhance diagnostic accuracy and reduce healthcare service times. Our approach includes employing artificial techniques to improve image quality for more effective model outcomes. Diagnosing skin diseases is visually difficult without dermatological expertise, as some skin lesions may result from other diseases. We utilized a public database for testing image enhancement which contains direct skin disease labels and available patient metadata (when present) for nuanced model predictions. Enhancement methods like ExcNet, CLAHE, and median filter addressed issues such as poor framing and inadequate lighting. The models tested included advanced neural networks such as CNN, ResNet-50, and MobileNet, among others. In our setting, low-light conditions were simulated via OpenCV darkening, and results should be interpreted accordingly. Within these constraints, the ExcNet deep learning technique showed the highest validation/test performance, with test accuracy reaching up to 91% in our experiments, while acknowledging potential label noise and dataset scope limitations.

Technology played a crucial role in combating the COVID pandemic, both in the rapid development of vaccines and the early detection of the virus. Consequently, numerous studies in the medical field have focused on leveraging the power of artificial intelligence for COVID-19 detection. However, in the medical domain, it is essential to have a clear understanding of the processes and algorithms used in decision-making, as these directly impact people’s health. Therefore, efforts have been made to implement explainable artificial intelligence techniques, enabling humans to understand and explain the deep learning algorithms used in disease detection. In this work, we present a novel approach to detecting COVID-19 in chest X-rays using a Variational Autoencoder model to identify lung anomalies. This methodology aims to highlight critical areas of the image, allowing healthcare professionals to identify them more effectively. Additionally, it seeks to provide clearer explanations of the decisions made by the artificial i ntelligence, r educing the complexity of the "black boxes" generated by deep learning neural networks. With this methodology, we hope to improve the effectiveness and reliability of early COVID-19 detection through chest X-rays.

When a company decides to use a chatbot, it must evaluate enough items to be able to take the right decision as to which one you should use to satisfy all your needs. There are times in which this decision is biased by trying to use the latest technologies available in the market and with this the full potential of the available products is not used. This project seeks to carry out research to implement an interaction chatbot basic text that allows it to be scalable mixed with other technologies such as ChatGPT and RPA (robotic process automation) can enhance its use to provide answers on unlimited topics added to having the possibility of carrying out tasks through RPA, which is all technology-oriented to the use of software with the aim of reducing human intervention in the use of applications computing, especially in repetitive tasks that vary very little in each iteration. RPA is a relatively new term, but not the concept behind it, since for decades it has been have looked for ways to optimize resources through software that performs specific functions in record time.

Recognizing variable stars is a task of interest in the astronomy community. Currently, this task has taken advantage of deep learning algorithms. However, these algorithms require a large amount of data to achieve high levels of precision. In this work, self-supervised learning is proposed to improve the classification of variable stars considering a reduced amount of data using recurrent networks. The experiments in Gaia dataset show that the proposed approach allows to improve performance, when compared with traditional initialization schemes, up to 7% and 13% in real databases in semi-supervised learning scenarios. In future work, we propose considering experiments with other variable star databases.

Marketing is essential for the success of any company today, both to project itself abroad and to achieve its business objectives. The personalized approach to marketing plays a crucial role, enabling more effective interaction with potential customers and reducing errors in promotional campaigns. One way to improve personalized marketing is by using artificial intelligence to predict the effectiveness of campaigns in converting users to customers. In this study, the use of Siamese neural networks for elevation modelling in a Chilean online retail company is proposed. The results obtained using this model are presented, as well as the classical elevation modelling techniques. These results show that the Siamese neural network model achieves an increase of more than 30% compared to the area under the Qini curve, while also being competitive in other metrics. As future work, it is planned to obtain more real data to carry out a better validation of this proposal as well as to compare the model with other proposed neural networks.

In the last decade, a large amount of data from vehicle location sensors has been generated due to the massification of GPS systems to track them. This is because these sensors usually include multiple variables such as position, speed, angular position of the vehicle, etc., and, furthermore, they are also usually recorded in very short time intervals. On the other hand, routes are often generated so that they do not correspond to reality, due to artifacts such as buildings, bridges, or sensor failures and where, due to the large amount of data, visual analysis of human expert is unable to detect genuinely anomalous routes. The presence of such abnormalities can lead to faulty sensors being detected which may allow sensor replacement to reliably track the vehicle. However, given the reliability of the available sensors, there are very few examples of such anomalies, which can make it difficult to apply supervised learning techniques. In this work we propose the use of unsupervised deep neural network models based on stacked autoencoders to detect anomalous routes in vehicles within Santiago de Chile. The results show that the proposed model is capable of effectively detecting anomalous paths in real data considering validation given by an expert user, reaching a performance of 82.1% on average. As future work, we propose to incorporate the use of Long Short-Term Memory (LSTM) and attention-based networks in order to improve the detection of anomalous trajectories.

Customer service is currently of great importance because this process supports the management of customer queries, requests and disagreements. In customer service, noncompliance with resolution times as well as closing times appearsas a recurring problem, which has a direct relationship with the customer experience. An alternative to improve this process isthrough the use of natural language processing algorithms to facilitate the complaint of customers considering platforms and solutions already applied in the market. In this paper, a customer complaint classifier is proposed using Google's Dialog Flow system in a Chilean company considering the implementation of a prototype composed of a text classification model and a system that manages customer interactions. This system allows an automatic categorization of customer interactions, allowing the tasks of a customer service representative to be reduced, minimizing the risks of misclassified interactions and avoiding re-categorizations that increase resolutiontimes. The proposed model allows improvements over 50% accuracy over the classification given by the pre-existing system based on user feedback. As future work we propose to create a chatbot system that allows improving user feedback.

Currently, air pollution is a highly important issue in society due to its harmful effects on human health and the environment. The prediction of pollutant concentrations in Santiago de Chile is typically based on statistical methods or classical neural networks. Existing methods often assume that historical values are known at a fixed geographic point, such that air pollution can be predicted at a future hour using time series analysis. However, these methods are inapplicable when it is necessary to know the pollutant concentrations at every point of the space. This work proposes a method that addresses the space-time prediction of PM2.5 concentration in Santiago de Chile at any spatial points through the use of the LSTM recurrent network model. In particular, by considering historical values of air pollutants (PM2.5, PM10 and nitrogen dioxide) and meteorological variables (temperature, wind speed and direction and relative humidity), measured at fixed monitoring stations, the proposed model can predict PM2.5 concentrations for the next 24 h in a new location where measurements are not available. This work describes the experiments carried out, with particular emphasis on the pre-processing step, which constitutes an important factor for obtaining relatively good results. The proposed multilayer LSTM model obtained R2 values equal to 0.74 and 0.38 in seven stations when considering forecasts of 1 and 24 h, respectively. As future work, we plan to include more input variables in the proposed model and to use attention-based networks.

Currently, cybersecurity has become one of the most important issues in the society, due to the growing technological evolution and use of digital platforms by organizations to connect with their users. A non-invasive biometric way of accessing an organizational platform is by verifying the presence of the user's face given by selfie photography in a database of authorized users. However, this procedure requires the prior construction of a database of authorized users, which prevents online verification of a person’s identity. A feasible way to carry out an online verification is by comparing the photos of the identity document and the person’s selfie. A system that allows this verification will improve the security of online use of a platform that prevents fraud and identity theft. In this work, we propose a deep neural network that allows verifying the identity of a person considering as inputs the photographs of an identity document and a selfie. This document presents experiments with various neural networks considering a public database of real camera photographs and identity documents. The results show that the use of a deep neural network with an ArcFace loss function configured with the database images of ID photos achieved a recognition rate of over 94%. Additionally, we test this procedure in a small validation sample of Chilean people obtaining similar rate. As future work, we propose the use of larger databases based on Chilean document data.

Since the beginning of 2020, the diagnosis of the COVID-19 virus has been a major problem that has affected the lives of millions of people around the world. The detection time for COVID-19 with a standard detection method ranges from approximately 1 to 5 days. An efficient and fast way to detect the presence of both the COVID-19 virus is through the use of artificial intelligence (AI) techniques applied to images obtained by lung radiography. Typically, AI algorithms to detect COVID-19 consider the whole picture. However, there may be parts that affect the performance of the classifier. Furthermore, these algorithms do not indicate which is the most relevant area of this disease. In this work, we propose a deep learning approach to detect the presence of COVID-19 in lung images by recognizing the most relevant areas affected by the virus without considering human supervision. In the experiment, we considered different proposals, where the best one obtained an 88% reduction of the logit loss with respect to the baseline based on random regions near the center of the image.

Marketing is a strategy that every company must implement today within its global plan both due to the need for external projection and for the achievement of commercial objectives. Currently, personalized marketing is key to the development of a company since it allows a better interaction with potential customers and the margin of loss or error in the direction of promotional campaigns is greatly reduced. One possibility to improve personalized marketing is the prediction of the effectiveness of campaigns to transform users into customers using artificial intelligence, so it is necessary to develop models that allow identifying profiles or segments of people who are more willing to answer positively to a campaign. This task corresponds to uplift modeling that predicts the incremental impact of the application of treatments on a population, which is typically performed using classical models such as the one-model approach, the class transformation approach, and the two-model approach. In this work, the use of multilayer neural networks is proposed to perform uplift modeling in a Chilean online retail company. The results of the proposed model as well as classical uplift modeling techniques are presented. These results indicate that the neuronal model allows an increase of more than 30 % in relation to the area of the Qini curve, while it is competitive in other metrics. As future work, it is planned to model a Siamese neural network with the cost function of uplift modeling directly.

Automatic recognition of visual objects using a deep learning approach has been successfully applied to multiple areas. However, deep learning techniques require a large amount of labeled data, which is usually expensive to obtain. An alternative is to use semi-supervised models, such as co-training, where multiple complementary views are combined using a small amount of labeled data. A simple way to associate views to visual objects is through the application of a degree of rotation or a type of filter. In this work, we propose a co-training model for visual object recognition using deep neural networks by adding layers of self-supervised neural networks as intermediate inputs to the views, where the views are diversified through the cross-entropy regularization of their outputs. Since the model merges the concepts of co-training and self-supervised learning by considering the differentiation of outputs, we called it Differential Self-Supervised Co-Training (DSSCo-Training). This paper presents some experiments using the DSSCo-Training model to well-known image datasets such as MNIST, CIFAR-100, and SVHN. The results indicate that the proposed model is competitive with the state-of-art models and shows an average relative improvement of 5% in accuracy for several datasets, despite its greater simplicity with respect to more recent approaches.

The kinesiological treatment of hand typically requires exercise either of mobility, elasticity or reflexes in the hands. This exercises requires to be evaluated by an expert, however, he could not be present. An alternative is the use of technologies, where eventually these exercises could be remotely modelled using 3D models for a more detailed representation of action respect to typical video records. Nonetheless, the hands are a very complex object as it is conformed by multiple parts, which difficults the application of typical sensors as RGB cameras. In this work, we propose to develop a software capable of implementing hand tracking system over a real-time infrared device. This system associates the sensor measures to precise position and pose of the hand. Besides, the software graph the hands three-dimensionally in a virtual environment considering different exercises that the user might perform. We expect that this proposal will be the basis of an application for kinesiological exercise in the hands, since this implementation technology also aims to increase the motivation of users to carry out the exercises and consequently, this could even mean a decrease in the recovery time.

Abstract In recent years, the pollution problem has gained great importance due to its socioeconomic implications for people regarding health or logistic issues. The pollution level classically is measured with specialized expensive detectors located in some few locations. In the case of Temuco city there are three such centralized pollution monitoring stations. An alternative approach for measuring the pollution level of cities makes use of inexpensive pollution sensors located on public transportation vehicles. Nonetheless, a drawback of this approach is that these inexpensive sensors can be sensitive to noise, vehicle movement, human intervention or technical failures. Therefore, it is relevant to be able to automatically detect inaccurate or failing sensors as they are multiple and cannot be submitted frequently to a technical revision. In this work, we propose a method to automatically detect these anomalous sensors by an unsupervised deep learning approach using autoencoders. This work is part of an ongoing project where massive data are not still available. In this context, the simulated output of mobile pollution sensors is generated by a time series model that systematically inserts outlier measurements. Our results indicate that the proposed detection method is able to reliably reproduce the data generated and to detect the simulated outliers with an accuracy of over 95%. A post-publication change was made to this article on 3 Jul 2020 to correct an author name.

Currently, a large amount of data is generated in the telemetry sector of vehicles in cities due to the continuous monitoring of vehicle trajectories through multiple sensors. Some trajectories generated by the sensors turn out not to correspond to the reality due to artefacts such as buildings, bridges or sensor failures, and where due to their large volume a manual verification of their correctness is not feasible. In this work, we propose the use of deep neural network models without supervision based on stacked autoencoders to detect atypical trajectories in vehicles within Santiago, Chile. The results show that the proposed model shows that it is able to detect that the atypical vehicle paths detected are at least 85% correct when considering the validation of a human expert. As future work, we propose to incorporate the use of LSTM networks in our model.

Today, there is growing interest in the automatic classification of a variety of tasks, such as weather forecasting, product recommendations, intrusion detection, and people recognition. "Mixture-of-experts" is a well-known classification technique; it is a probabilistic model consisting of local expert classifiers weighted by a gate network that is typically based on softmax functions, combined with learnable complex patterns in data. In this scheme, one data point is influenced by only one expert; as a result, the training process can be misguided in real datasets for which complex data need to be explained by multiple experts. In this work, we propose a variant of the regular mixture-of-experts model. In the proposed model, the cost classification is penalized by the Shannon entropy of the gating network in order to avoid a "winner-takes-all" output for the gating network. Experiments show the advantage of our approach using several real datasets, with improvements in mean accuracy of 3-6% in some datasets. In future work, we plan to embed feature selection into this model.

Within the increasing automation of tasks, it is necessary to obtain relevant information from images, to have clarity of the actions that must be performed. In this process, it is possible to detect objects individually, however this mode can generate a considerable error due to the enormous number of ways in which an object can be presented. It is therefore necessary to improve the level of accuracy, and a way to achieve this is taking into account the relationships between objects as well. Until now, context-oriented models generate relationships between objects without discriminating between input images. It is necessary that the model operates on each particular image, to reduce the error and obtain a conclusive result. In response to the above, this work proposes an adaptive probabilistic context-model. The model in question is a functional that processes the occurrence probability of the objects in each image, and generates the relations between the detectable objects using a Bayesian network in the form of a tree. This model is updated with the features of each image, requiring a minimization of the quadratic error through a numerical approximation, obtained by Newton-Raphson method. Comparisons were made between different heuristic proposals, as well as tests on different context-oriented databases, in order to validate the results. On the other hand, tests were performed using features extracted through histograms of orient gradients and Deep Learning. It was found that an improvement in the prediction of the order of 20% is feasible in the testing process.

Today, organizations are beginning to realize the importance of using as much data as possible for decision-making in their strategy. The finding of relevant patterns in enormous amount of data requires automatic machine learning algorithms, among them, a popular option is the mixture-of-experts that allows to model data using a set of local experts. The problem of using typical learning algorithms over Big Data is the handling of these large datasets in primary memory. In this paper, we propose a methodology to learn a mixture-of-experts in a distributed way using PETUUM platform. Particularly, we propose to learn the parameters of mixture-of-experts by adapting the standard stochastic gradient descent in a distributed way. This methodology is applied to people detection with standard real datasets considering accuracy and precision metrics among other. The results show a consistent performance of mixture-of-experts models where the best number of experts varies according to the particular dataset. We also evidence the advantages of the distributed approach by showing the almost linear decreasing of average training time according to the number of processors. In a future work, we expect to apply this methodology to mixture-of-experts with embedded variable selection.

In these days, there are a growing interest in pattern recognition for tasks as prediction of weather events, recommendation of the best route, intrusion detection or face detection. Each of these tasks can be modelled as classification problem, where a common alternative is to use an ensemble model of classification. A well-known example is given by Mixture-of-Experts model, which represents a probabilistic artificial neural network consisting of local experts classifiers weighted by a gate network, and whose combination creates an environment of competition among experts seeking to obtain patterns of the data source. We observe that this architecture assume that one gate influence only one data point, consequently the training can be misguided in real datasets where the data is better explained by multiple experts. In this work, we present a variant of regular Mixture-of-Experts model, which consists of maximizing of the entropy of gate network in addition to classification cost minimization. The results show the advantage of our approach in multiple datasets in terms of accuracy metric. As a future work, we plan to apply this idea to the Mixture-of-Experts with embedded feature selection.

Pedestrian detection has multiple applications as video surveillance, automatic driver-assistance systems in vehicles or visual control of access. This task is challenging due to presence of factors such as poor lighting, occlusion or uncertainty in the environment. Deep learning has reached many state-of-art results in visual recognition, where one popular and simple variant is stacked autoencoders. Nonetheless, it is not clear what is the effect of each stacked autoencoders parameter in pedestrian detection performance. In this work, we propose to revise the feature representation for pedestrian detection considering the use of deep learning using stacked autoencoders with a sensitivity analysis of relevant parameters. Additionally, this paper presents a methodology for feature extraction using stacked autoencoders. The experiments show that this model is capable of creating a meaningful visual descriptor for pedestrian detection, which improves the detection performance in comparison to baseline techniques without an optimal setting of parameters. In presence of occlusion or poor people images, we found diffuse and distorted visual patterns. A future avenue is the learning of the degree of noise for improving the generalization capabilities of the learned features.

The high rate of university dropout and low graduation rates are very relevant social problems today. Since there are many possible causes of desertion and university graduation, in this paper, we propose to find, analyze and weigh the factors that allow predicting if a student will drop out or graduate according to prior information available using data mining techniques and statistical models. We will focus in the case of Catholic University of Temuco, using real data from that institution. This study reveals relevant variables in opinion of human experts, which demonstrates the ability of automatic models to represent the dropout and graduation at the university.