In this work, we propose an approach for unsupervised depth and
camera motion estimation. We focus on two aspects of the view synthesis
approach. We learn attention maps that are used to increase the error
contribution of the meaningful regions on the error computation between
the synthesized and the real views. Moreover, we propose to approximate
the uncertainty of pixels of being visible by mapping the depth consistency
to probabilities using a Gaussian function. The attention maps are learned
simultaneously with the depth network in an end-to-end manner.
We evaluated our method on the depth estimation and odometry tasks
of the KITTI benchmark. The inclusion of the attention and visibility
maps shows an improvement in depth and camera motion estimation tasks
when compared with a competitive baseline method.

Learning Attention Maps for Self-supervisedDepth and Camera Motion Estimation in Monocular Videos.

The proposed technique is based on Mask-RCNN and was evaluated using the INCAN Database. We obtained a ganglion and spiculated mass detection rate of 87% and segmentation performances with 82.9% sensitivity, 93.7% precision, and 88.0 F1. The database is validated by 5 experts, an oncologist specializing in breast pathology, three radiation oncologists and a mastologist. This approach allows us to employ deep learning techniques to provide assistance to healthcare professionals in the medical diagnosis for breast cancer.

Automatic classification of the ganglion and spiculated mass in mammography

This poster explains a novel document representation presented at NAACL 2018 in NOLA. The representation is  called  Multi-Resolution  Represen-tation (MulR), an aims to improve the early detection of  risks  in  social  media  sources.   The  goal is to effectively identify the potential risk using  as  little  evidence  as  possible  and  with as  much  anticipation  as  possible.   MulR  allows  us  to  generate  multiple  “views”  of  the text.   These views capture different semantic meanings for words and documents at different  levels  of  granularity,  which  is  very  useful  in  early  scenarios  to  model  the  variable amounts of evidence.  The experimental evaluation shows that MulR using low resolution is better suited for modeling short documents (very early stages),  whereas large documents (medium/late stages) are better modeled with higher resolutions.  We evaluate the proposed ideas in two different tasks where anticipation is critical: sexual predator detection and depression detection.  The experimental evaluation for these early tasks revealed that the proposed approach outperforms previous methodologies by a considerable margin. In this poster we will also explain some of our ongoing work related with the idea of concept representation applied to sentiment analysis and other text classification problems.

Early Text Classification using Multi-Resolution Concept Representations

The smart meters allow to monitor the data of consumption/production of electric energy in certain periods of time and sending them to data servers for later billing. The objective of this work is to analyze the data directly from the device (edge ​​computing) to determine if there are anomalous consumption/production patterns that allow determining if it is a possible theft of energy by the users or the electric company. To achieve this, the smart meter predicts through its historical data the following consumption value and based on previous and future estimates, it deduces if it is a possible energy theft or alteration of the production  The technique used is a variant of the ARIMA model in time series. The results show that the proposed model can predict with an acceptable margin of error a possible attempt to energy theft or tampering data production.

 Data Analytics in Smart Meters for the Detection of Energy Theft

Social media has become one of the principal sources of news consumption, one of their properties is the speed in which the content is created and spread, but the content is not always verified by the users before they share them. This have made easy to generate intentionally false content, there exists different approaches to tackle the detection of fake news mostly of which use English texts for the analysis. We take these works as a basis to analyze and propose some attributes useful in the detection of fake news in Spanish, there exists differences in the way in which the fake news content is generated between language because of the cultural differences and the target audience. We make use of properties of the texts as readability metrics to analyze the difference between content generated by professional journals and not recognized webs, and as a social media approach we use Twitter to analyze how the content is spread between users of this social network.

Detection of Fake News based on readability

                                   Prediction of production in Stevia plants using artificial neural networks
                                                          A. F. Chávez-Villanueva, Alvizar Manzo R.
                                                        Instituto Tecnológico Superior de los Reyes 

The objective of this research is to evaluate different types of artificial neural networks, which allow to understand, analyze and predict the response of Stevia plants, at different concentrations of nutrients in order to increase production. To carry out the aforementioned, a database with photographs of different leaves of each Stevia plant for 3 months was generated. These data were processed in order to implement them as input vectors in the network. Subsequently, from each leaf, the area was calculated for its implementation as objective vector. Once the network was trained to process the data, the respective production predictions were generated in Stevia plants, for later experimental verification.   

Keywords: Stevia, plants, artificial neural networks, analyze, predict, increase, production

Prediction of production in Stevia plants using artificial neural networks

Currently, the exponential growth of large cities has drastically reduced the amount of healthy green areas recommended by public health institutions. Adding the fact of having to monitor huge areas like the 1495 km2 of Mexico City, the problem is difficult to address. With the aim of achieving a viable solution, this work presents a software capable of automatically classifying the quality of the terrain with 72% accuracy from images of a drone that flies 30m high, which allows analyzing small details of the terrain. The software and the dataset will be available for anyone's use and allows the possibility of retraining the designed deep network.


A novel software for analyzing the natural landscape surface using deep learning

World Health Organisation (WHO) estimates around 400 million cases of rare diseases worldwide and 5000 to 8000 different varieties. Most of them have genetic bases, which hinders medical diagnosis. These conditions usually occur as a consequence of low frequency single nucleotide polymorphism (SNPs). For this reason, genomic studies have had a great impact in accelerating the aforementioned diagnosis. However, common bioinformatics analysis requiere an unfeasible amount of working hours to cover the whole human genome. Our research uses a Machine Learning approach to address this problem.
In-silico work aims to classify variants in five groups: benign, likely benign, uncertain significance, likely pathogenic and pathogenic. For coding regions of the genome, which represents less that 2% of total DNA, the typical approach works perfectly well. Yet, there is no established protocol on how to identify disease-causing SNPs in long no-coding regions. If successful, our approach will help prioritize variants from patients suffering from undiagnosed rare diseases.
For doing so we make use of manually curated public databases. Here each researcher adds variants specifying its chromosome, location, reference allele, alternate allele and tagged with one of the above mentioned clinical significances. This information is used to annotate each variant with a number of different scores based on its characteristics. Most of these scores use amino acidic properties to determine their pathogenicity, making no-coding variants hard to characterise. Therefore, a particularly precise curation of the dataset is required to avoid biases in the biological properties that determine the pathogenicity of no-coding SNPs.
Given our data, the classification process will be divided in two parallel workflows, each one involving two step. Firstly, due to their difference in features, no-coding variants will be separated from protein coding variants. Both datasets will continue the process in parallel. Secondly, a categorical classifier will determine whether each variant has an uncertain significance or not. And thirdly, those variants with some clinical significance will be classified using a continuous algorithm in order to predict their “degree” of pathogenicity.
Once the most adequate classifiers are determined, the same workflow (annotation + variant prioritization) will be applied to variants obtained from rare diseases patients. We hope this approach will help us determine the genetic bases of these pathologies, helping diagnose patients in a delicate situation.

Machine Learning approach to identify pathogenic variants in non-coding regions of the human genome.

El objetivo de este trabajo es crear una metodología para la generación de potenciales de red neuronal aplicables a las diferentes fases de la alúmina, mediante la cual sea posible calcular diferentes propiedades comparables y con la misma precisión a las obtenidas a través de los cálculos de referencia de DFT, por ejemplo, la predicción de geometrías, energías, fuerzas atómicas y cargas atómicas, empleando un costo computacional más bajo, con un mayor número de átomos (hasta millones de átomos) y con cálculos de Dinámica Molecular (MD) de hasta nanosegundos, para predecir propiedades que se asemejan a lo que sucede en el laboratorio.
Los resultados a la fecha son la generación de un potencial de Red Neuronal Artificial de la α-Al2O3 a partir de los cálculos de Dinámica Molecular de primeros principios por el método de DFT, de los cuales se alimentó a la red neuronal 2820 estructuras, en donde se tomaron como variables los ambientes químicos locales, es decir, sus primeros y segundos vecinos, obteniendo un potencial de red neuronal (RN), en el cual se evaluó cual es su mejor arquitectura, es decir, cuántos nodos en sus capas ocultas es el óptimo, evaluando el tiempo y el grado de ajuste, observado que la mejor arquitectura para el alúmina es de 70-10-10-1, que es igual a 70 nodos en su capa de entrada, 10 en sus dos capas ocultas y 1 en su capa de salida. Anteriormente se había realizado un potencial en el cual se utilizaron 9 mil muestras, sin embargo, estas no tuvieron un buen ajuste debido a que los datos de entrada de la red neuronal no eran los adecuados para este tipo de estructuras, el potencial de predicción variaba entre 5-100 meV/atomo, así que se optó por solo utilizar los cálculos de la dinámica molecular a diferentes temperaturas. Se realizaron evaluaciones al potencial de red neuronal, ya que solo se había visto si funcionaba el potencial de RN de forma visual en la dinámica molecular; estas evaluaciones constaron de las simulaciones a diferentes tiempos de la ecuación de movimiento y con estructuras que no se alimentaron a la red neuronal, para verificar que el potencial fuera continuo y de un perfil suave, en esta evaluación se observó que no hay variación abrupta de las energías por  lo que se puede decir que el potencial tiene un perfil de energía continuo.  Adicionalmente, se realizaron cálculos de energía para diferentes tamaños de estructura, desde 30 hasta 1 millón de átomos, empleando para tal fin computadoras comerciales de escritorio, a través de tiempos con una tendencia lineal, ésto para poder evaluar el escalamiento del potencial de RN. Se realizó una simulación de Dinámica Molecular de 100,000 fs de una supercelda 2x2x2, a 1000 K y un pasos de 1 fs. De esta simulación, cada 1000 pasos se tomaron las coordenadas atómicas y las energías y se recalcularon por el método de DFT, en donde se observó que este potencial tiene un ajuste de 1 meV/átomo, en comparación con los cálculos de primeros principios. 


Análisis estructural de las Alúmina usando DFT e Inteligencia Artificial

Manual image segmentation of ecographies was done In this work tumor classification was performed support vector machine (SVM) algorithm was used to classify benign and malign tumors.94% of echographies were correctly classified.

Tumor detection and classification using ultrasound images

Experimental psychology is an established field whose main interest is the study of psychological phenomena by empirical means. One of those phenomena is the perception of time in the seconds to minutes range, which has attracted a great deal of research, and several mathematical models have been proposed. Most current studies have been done in the framework of one theory or other, and their data is often summarized in a few measures and then analyzed by classical statistical methods. We think that the richness of the data can be exploted more with the help of other approaches, like those in the machine learning literature. Here we present an experiment of time perception using a procedure on which the animals (rats) must learn when to respond in order to obtain a reward, and then we alter the rats uncertainty of this trained interval by inserting a delay beween the rats responses and the delivery of the reward. We applied some ML techniques to gain insights about how animals learn when to start and stop responding, and how the uncertainty affects their estimation of the trained interval. Among the problems explored are density ratio estimation, distributional change detection, mutual information estimation and change-point detection. 

Machine learning applications in experimental psychology: a study of the perception of time

El sector energético Mexicano ha experimentado cambios sin precedentes: las reformas legislativas del año 2013 abrieron, por primera vez en 70 años, el mercado de hidrocarburos de México a empresas privadas nacionales y extranjeras, la penetración de energía renovable en el país ha crecido rápidamente, sin embargo las importaciones de gas natural y derivados de petróleo aumentan la dependencia comercial de estos energéticos con EEUU,  mientras se discuten la viabilidad de grandes proyectos de inversión pública y privada. Esto ha dado como resultado un entorno de decisión extremadamente complejo que exacerba los desafíos tradicionales asociados con el desarrollo de sistemas de infraestructura resilientes en las naciones emergentes. 
Utilizamos el marco de trabajo de la Toma de Decisiones Robustas (RDM) para ayudar a los tomadores de decisiones en el diseño de estrategias de inversión y/u operación para cimentar el desarrollo de una infraestructura de almacenamiento y transporte de combustibles resiliente. Primero, desarrollamos un modelo de optimización que simula la distribución de petrolíferos en México, sujeto a restricciones. Adicionalmente incorporamos la incertidumbre de variables clave a través de un proceso exploratorio computacional. Finalmente hacemos un análisis de clasificación de escenarios con técnicas de aprendizaje de máquina para determinar nodos vulnerables, redundancias de infraestructura y costos necesarios para satisfacer la demanda bajo diferentes estresores que nos permitan determinar configuraciones de red resilientes.


Resiliencia en sistemas de infraestructura complejos bajo condiciones de incertidumbre profunda: un análisis de la infraestructura de almacenamiento y transporte de combustibles de México

The following document has as objective to elaborate a model of time series that allows to efficiently carry out the forecast of sales using the STL model. The reason for this document, therefore, is to illustrate the appropriate process to elaborate an STL model using an optimal value of Span, and to compare the results obtained from the residuals of the model, and in turn, with the result of the final RSME. .

Sales Forecast using a  Machine Learning Time Series model.

Catalysis is an interdisciplinary and complex field where several pieces of information must be put together to design a successful working catalyst. In recent years, theoreticians have contributed to accelerate the discovery of new catalytic materials by putting together information repositories like «Catalysis-Hub», but often the models address either the molecular or the engineering aspect of the reaction. Until now, the comparison in preparation methods is difficult even with materials prepared in the same laboratory. In this work, synthesis parameters for NiMo sulfide catalysts were extracted from existing literature adding specific algorithms to the ChemDataExtractor tool. It was found that natural language processing techniques can be used to extract information and gain knowledge from a great number of systems and allow to find hidden or misregarded links between preparation conditions. 

Machine learning and natural language processing for the identification of synthesis parameters of NiMo sulfide catalysts 

Inspiration in nature has been widely explored, from macro to micro-scale. When looking into chemical phenomena, stability and organization are two properties that emerge. Recently, artificial hydrocarbon networks (AHN), a supervised learning method inspired in the inner structures and mechanisms of chemical compounds, have been proposed by as a data-driven approach in artificial intelligence. AHN have been successfully applied in data-driven approaches, such as: regression and classification models, control systems, signal processing, and robotics. To do so, molecules –the basic units of information in AHN– play an important role in the stability, organization and interpretability of this method. Interpretability, saving computing resources, and predictability have been handled by AHN, as any other machine learning model. This short paper aims to highlight the challenges, issues and trends of artificial hydrocarbon networks as a data-driven method. Throughout this document, it presents a description of the main insights of AHN and the efforts to tackle interpretability and training acceleration. Potential applications and future trends on AHN are also discussed.

RIIAA Submissions