Noticimat 41

Ponente: Ranganath Krishnan (Intel Labs)

Título: Uncertainty Calibration and Robustness in Deep Learning

Hora: 12:30 GMT-6

Enlace de transmisión: https://bluejeans.com/932081805

Resumen: Obtaining reliable and accurate uncertainty estimates from deep neural networks is essential to build trustable and robust AI systems. A well-calibrated model should be accurate when it is certain about its prediction and indicate high uncertainty when it is likely to be inaccurate. Uncertainty calibration is a challenging problem as there is no ground truth available for uncertainty estimates. In this talk, I will present an overview of practical methods for uncertainty quantification in deep learning and challenges of model calibration under distributional shift. I will focus on our novel uncertainty calibration methods that leverage the relationship between accuracy and uncertainty to obtain well-calibrated uncertainties in probabilistic deep neural networks, along with the empirical results on robustness to distributional shift (dataset shift and out-of-distribution data). I will also discuss our recent work on scaling variational inference in large-scale Bayesian deep neural networks with Empirical Bayes, illustrating with real-world applications.

 

Ponente: Dra. Claudia Esteves Jaramillo (UG)

Título: Representación de datos para análisis de movimiento humano

Hora: 12:30 GMT-6

Enlace de transmisión: https://bluejeans.com/830896280

Resumen: En esta charla les contaré cómo adquirimos y representamos datos para hacer análisis y generación automática de movimiento humano. Describiré el equipo que tenemos en el laboratorio y los tratamientos necesarios para poder modelar los datos que de aqui salen. Al final contaré algunas de las aplicaciones en las que estamos trabajando.

 

Enlace de transmisión: http://vbac.wikidot.com

 

Hora: 15:00 GMT-6

Ponente 1: Simon Donaldson (Fields Medal)

Título: K-stability and scalar curvature

Resumen: This will be an overview talk about existence results in complex differential geometry connected to the notion of K-stability. We will explain the analogies with the corresponding results, going back to Narasimhan and Seshadri, for holomorphic vector bundles and outline some strategies of proofs that have been employed. We will illustrate the general with a discussion of the case of toric manifolds.  

 

Hora: 16:00 GMT-6

Ponente 2: Chenyang Xu

Título: An algebraic construction of K-moduli space

Resumen: K-stability of Fano varieties has become a fast developed topic in algebraic geometry. One major output is the construction of moduli spaces of K-(semi,poly)-stable Fano varieties, which resolves a number of pathological issues for families of general Fano varieties. The purely algebraic construction is built on a systematical study of K-stability using higher dimensional geometry, including a more comprehensive understanding of the notion of K-stability (for Fano varieties).  

 

Ponente: Scott Mullane (Institute of Mathematics; Goethe University)

Título: The Hurwitz Picard Rank Conjecture for d>g-1

Hora: 15:30 GMT-6

Enlace de transmisión: https://bluejeans.com/759804451

Resumen: Despite the utility of Hurwitz spaces in the study of a number of the birational aspects of the moduli space of curves, many open questions on Hurwitz spaces persist. I'll show how the perspective of the exact differentials on curves can be used to prove, that as conjectured, the rational Picard group of the moduli space of simply branched degree d covers of the rational line by smooth genus g curves is trivial for d>g-1. Further, I'll discuss how this perspective yields results on open questions on the irreducibility of non-simple Hurwitz spaces and has applications to the birational geometry of moduli spaces of pointed rational curves.

 

Ponente: Diego Corro (Instituto de Matemáticas-UNAM) 

Título: Variedades con foliaciones asféricas

Hora: 16:30 GMT-6

Enlace de transmisión: https://bluejeans.com/560153998 

Resumen: En esta plática presentare un tipo particular de foliaciones singulares Riemannianas, es decir foliaciones singulares compatibles con una métrica Riemanniana. Dada una foliación singular Riemanniana $(M,mathcal{F})$ cuyas hojas son homeomorfas a toros, daremos una serie de invariantes que determinan la vecindad tubular de una hoja.

En el caso en que $M$ es simplemente conexa, y existe para la proyección $pi colon M o  M/mathcal{F}$ una sección, i.e. una función inversa derecha $sigma colon M/ mathcal{F} o M$ de $pi$, veremos que estos invariantes determinan a la foliación, salvo homeomorfismo foliado.

En el caso en que la foliación tiene codimensión $2$, se puede dar un argumento para demostrar que las foliaciones asféricas en variedades simplemente conexas están dadas por acciones suaves del toro.

 

Ponente: Dr. Carlos E. Rubio-Medrano (Texas A&MUniversity-Corpus Christi)

Título: Cybersecurity Perspectives on Mobile Augmented Reality: The Coolest Emerging Technology Just Around the Corner

Hora: 11:00 GMT-6

Enlace de transmisión: https://bluejeans.com/138086631

Resumen: Mobile Augmented Reality (MAR) has progressed significantly in the last few years. As an example, Pokémon GO, a popular MAR application, turned the world upside down in 2016 by allowing for gamers to actively roam their surroundings using their mobile phones. As of today, almost every major computing company has become interested in MAR: Google, Facebook, Apple, and others are expected to release MAR-enabled products in the next two years. Even car manufacturers have jumped on the bandwagon: Mercedes Benz will incorporate MAR features into the windshield of new cars starting in 2022.

In order to enhance the user experience, MAR applications are increasingly collecting sensitive data such as personal information, location coordinates, usage patterns, etc., which has been found to be problematic by the cybersecurity community. First, there are some issues regarding the privacy and the proper access mediation of such sensitive information. Also, the intrusive affectation of the user experience has also been identified as a major concern. As an example, malicious MAR content such as unwanted advertisements placed in physical spaces, which may not only affect the MAR experience, but may also represent a threat to the physical safety of users. Finally, the invasion of sensitive physical spaces by MAR applications has been also recognized as problematic. As an example, Pokémon GO has been banned from physical places such as hospitals, schools, and memorials worldwide.

During my talk, I will elaborate on the newest developments on MAR, as well as the cybersecurity issues just described. In addition, I will introduce Space Sensitive Access Control (SSAC), an ongoing research effort in collaboration with partners at Arizona State University (ASU), the University at Buffalo, and the Austrian Institute of Technology (AIT). SSAC is not only effective at counteracting the aforementioned problems, but it is also based on precise rules that regulate the interaction between users, MAR applications, and physical spaces, representing a convenient one-size-fits-all solution for handling different cases in practice, and serving a strong foundation for developing secure MAR applications.

 

Ponente: Dr. Adnan Darwiche (UCLA)

Moderador: Enrique Succar

Título: Razonamiento y aprendizaje manipulable con aritmética y circuitos booleanos

Hora: 12:00 GMT-6

Enlace de transmisión: https://youtu.be/S_e1DLp0w9A

Resumen: Los circuitos manipulables han estado jugando un papel cada vez más importante en IA y más allá, particularmente en las áreas de la simbología y la probabilística razonamiento. Esto incluye proporcionar un enfoque sistemático para abordando los problemas más allá de NP, permitiendo que uno aprenda de ciertos combinaciones de conocimientos y datos, y el razonamiento sobre la el comportamiento de algunos sistemas de aprendizaje de máquinas. En esta charla, el Dr. Darwiche revisará los fundamentos teóricos y algunas aplicaciones de los tractables los circuitos, tanto los booleanos como los aritméticos.

 

 

Ponente: Yoonkyung Lee (Ohio State University)

Título: Assessment of case influence in support vector machine

Hora: 15:00 GMT-6

Registro previo en http://cpfcomputacionmatematica.eventos.cimat.mx

Resumen: Support vector machine (SVM) is a very popular technique for classification. A key property of SVM is that its discriminant function depends only on a subset of data points called support vectors. This comes from the representation of the discriminant function as a linear combination of kernel functions associated with individual cases. Despite the direct relation between each case and the corresponding coefficient in the representation, the influence of cases and outliers on the classification rule has not been examined formally. Borrowing ideas from regression diagnostics, we define case influence measures for SVM and study how the classification rule changes as each case is perturbed. To measure case sensitivity, we introduce a weight parameter for each case and reduce the weight from one to zero to link the full data solution to the leave-one-out solution. We develop an efficient algorithm to generate case-weight adjusted solution paths for SVM. The solution paths and the resulting case influence graphs facilitate evaluation of the influence measures and allow us to examine the relation between the coefficients of individual cases in SVM and their influences comprehensively. We present numerical results to illustrate the benefit of this approach.

 

 

Ponente: Vikash K. Mansinghka (MIT)

Título: Probabilistic programming and artificial intelligence

Hora: 15:00 GMT-6

(Registro previo en http://cpfcomputacionmatematica.eventos.cimat.mx)

Resumen: Our minds unconsciously explore vast spaces of possible models for sparse data, often identifying probable explanations and useful actions in fractions of a second. The robustness, flexibility, speed, and data efficiency of human common sense --- and of human expert reasoning --- appears out of reach for purely symbolic, probabilistic, and differentiable / "neural" approaches to AI. How can we narrow these gaps between artificial and natural intelligence?

This talk will review progress towards solving these problems via probabilistic programming, an emerging field at the intersection of artificial intelligence, probability theory, and programming languages. It will introduce basic concepts, including generative programs and inference programs, via common-sense scene understanding examples written in Gen. Gen is a new general-purpose probabilistic programming platform developed at MIT. Unlike differentiable programming platforms such as TensorFlow and PyTorch, generative programs in Gen can freely combine symbolic, probabilistic, and differentiable components, and Gen inference programs can integrate exact inference with custom variational, optimization-based, and Monte Carlo approximations. Gen is being used by a growing community of researchers and engineers working in domains such as robotics, systems biology, climate science, cognitive science, and artificial intelligence. Gen is also central to a new project at MIT aimed at reverse engineering human 18-month-old common sense.

This talk will also show how to scale up applications via domain-specific probabilistic languages. Examples include automated Bayesian data modeling by Bayesian learning of generative program structure and parameters, and expert systems for deduplication and cleaning of real-world databases with millions of records.

 This talk is based on joint work with Marco Cusumano-Towner, Tan Zhi-Xuan, Feras Saad, Alex Lew, and other students and research staff from the MIT Probabilistic Computing Project.

Con el tema “Género y tecnología”, el próximo miércoles 25 de noviembre a las 15:00 GMT-6 se realizará la tercera mesa de diálogo Construyamos mediante el diálogo la IA en México. 

 

Irene Soria Guzmán (UAM-Xochimilco)

Juana Martínez (Women in data science)

Mónica Vázquez (IIMAS/ UNAM)

Pablo Aguilera (Mozilla)

Moderadora: Sofía Trejo (AIA)

 

Registro en https://docs.google.com/forms/d/e/1FAIpQLSfbaHP1bt5bcW8M1bsWh0lY3c8LAj6qClv8eGYyRpG73VwkTQ/viewform

 

Más información en https://www.consorcioia.mx/mesas-de-dialogo

 

El próximo jueves 26 de noviembre se realizará en línea la II Escuela de Modelación Estadística para Ciencias Sociales, dirigida a estudiantes de las carreras afines a la temática de la escuela de los últimos semestres, estudiantes de posgrado, investigadores y tomadores de decisiones, principalmente.

 

Registro de asistentes y mayores detalles en: http://emecs2020.eventos.cimat.mx

Felicitamos a los siguientes estudiantes:

 

Rogelio Cruz Romero, quien el miércoles pasado presentó su tesis Adaptive Batching First Order Optimization Methods For Machine Learning Applications y obtuvo el grado de Maestro en Ciencias con Especialidad en Computación y Matemáticas Industriales. En el jurado participaron el Dr. Arturo Hernández Aguirre (CIMAT), presidente; el Dr. Carlos Segura González (CIMAT), secretario; y el Dr. Óscar Susano Dalmau Cedeño (CIMAT), vocal y director de la tesis.

 

Víctor Manuel González Ruiz, quien el pasado jueves ovtuvo el grado de Maestro en Ciencias con Especialidad en Probabilidad y Estadística, con la presentación y defensa de su tesis Función Característica del Modelo Rugoso de Heston, la cual contó con el Dr. Ehyter Matías Martín González (UG) como lector especial y codirector. Los integrantes del jurado fueron la Dra. Ekaterina Todorova Kolkovska (CIMAT), presidente; el Dr. Harold Andrés Moreno Franco (HSE-Moscú), secretario; y el Dr. José Luis Ángel Pérez Garmendia (CIMAT), vocal y director de la tesis.

 

Richard Fabián Arteaga Ospina, quien este viernes obtuvo el grado de Maestro en Ciencias con Especialidad en Computación y Matemáticas Industriales, con la presentación y defensa de su tesis titulada Un Planificador SST Eficiente para Trayectorias Mínimas en Tiempo para un DDR con Dinámica de Segundo Orden entre Obstáculos, misma que tuvo la lectura especial y codirección del Dr. Israel Becerra Durán (CIMAT-CONACYT). En el jurado estuvieron el Dr. Eduardo Francisco Morales Manzanares (INAO-CIMAT), presidente; el Dr. Marco Antonio Morales Aguirre (ITAM), secretario; y el Dr. Rafael Eric Murrieta Cid (CIMAT), vocal y director de la tesis.

El lunes 23 y miércoles 25 de noviembre, se llevarán a cabo los Rallies Matemáticos en su modalidad Síncrona (en vivo a través de Zoom). El rally matemático representa una oportunidad para motivar a los estudiantes para acercarse al conocimiento y a la habilidad de resolución de problemas a través del juego. Además, se espera que sirva de motivación para que los alumnos se inscriban a las olimpiadas matemáticas. Los rallies contarán con la participación de 72 estudiantes de primaria y 72 estudiantes de secundaria del estado de Guanajuato.

 

Del lunes 23 al miércoles 25, se realizarán cuatro sesiones del taller "Curso de Círculos Matemáticos", dirigido a docentes de secundaria del estado de Guanajuato. Durante estas sesiones se espera que los participantes se familiaricen con la dinámica y el tipo de ejercicios de los Círculos Matemáticos y puedan replicar las actividades con sus estudiantes.