Carlos González-Gutiérrez

Comparative Study of Neural Network Frameworks for the Next Generation of Adaptive Optics Systems

Many of the next generation of adaptive optics systems on large and extremely large telescopes require tomographic techniques in order to correct for atmospheric turbulence over a large field of view. Multi-object adaptive optics is one such technique. In this paper, different implementations of a tomographic reconstructor based on a machine learning architecture named “CARMEN” are presented. Basic concepts of adaptive optics are introduced first, with a short explanation of three different control systems used on real telescopes and the sensors utilised. The operation of the reconstructor, along with the three neural network frameworks used, and the developed CUDA code are detailed. Changes to the size of the reconstructor influence the training and execution time of the neural network. The native CUDA code turns out to be the best choice for all the systems, although some of the other frameworks offer good performance under certain circumstances.

Using GPUs to Speed up a Tomographic Reconstructor Based on Machine Learning

The next generation of adaptive optics (AO) systems require tomographic techniques in order to correct for atmospheric turbulence along lines of sight separated from the guide stars. Multi-object adaptive optics (MOAO) is one such technique. Here we present an improved version of CARMEN, a tomographic reconstructor based on machine learning, using a dedicated neural network framework as Torch. We can observe a significant improvement on the training an execution times of the neural network, thanks to the use of the GPU.

Improving Adaptive Optics Reconstructions with a Deep Learning Approach

The use of techniques such as adaptive optics is mandatory when performing astronomical observation from ground based telescopes, due to the atmospheric turbulence effects. In the latest years, artificial intelligence methods were applied in this topic, with artificial neural networks becoming one of the reconstruction algorithms with better performance. These algorithms are developed to work with Shack-Hartmann wavefront sensors, which measures the turbulent profiles in terms of centroid coordinates of their subapertures and the algorithms calculate the correction over them. In this work is presented a Convolutional Neural Network (CNN) as an alternative, based on the idea of calculating the correction with all the information recorded by the Shack-Hartmann, for avoiding any possible loss of information. With the support of the Durham Adaptive optics Simulation Platform (DASP), simulations were performed for the training and posterior testing of the networks. This new CNN reconstructor is compared with the previous models of neural networks in tests varying the altitude of the turbulence layer and the strength of the turbulent profiles. The CNN reconstructor shows promising improvements in all the tested scenarios.

Compensating Atmospheric Turbulence with Convolutional Neural Networks for Defocused Pupil Image Wave-Front Sensors

Adaptive optics are techniques used for processing the spatial resolution of astronomical images taken from large ground-based telescopes. In this work are presented computational results from a modified curvature sensor, the Tomographic Pupil Image Wave-front Sensor (TPI-WFS), which measures the turbulence of the atmosphere, expressed in terms of an expansion over Zernike polynomials.

Outcome Prediction for Salivary Gland Cancer Using Multivariate Adaptative Regression Splines (MARS) and Self-Organizing Maps (SOM)

Over the last decades, advances in diagnosis and tissue microsurgical reconstruction of soft tissues have modified the therapeutic approach to salivary gland cancers, but long-term survival rates have increased only marginally. Due to the relatively low frequency of these tumors together with their diverse histopathological types, it is not easy to perform a prognosis assessment. Multivariate adaptative regression splines (MARS) is a data mining technique with a well-known ability to describe a response starting from a large number of predictors. In this work, MARS was used for determining the prognosis of cancers of salivary glands using clinical and histological variables, as well as molecular markers. Here, we have generated four different models combining different sets of variables, with sensitivities and specificities ranging from 95.45 to 100%. Specifically, one of these models which combined five clinical variables (Tumor size –T–, neck node metastasis –N–, distant metastasis –M–, age and number of tumor recurrences) plus one molecular factor (gelatinase B -MMP-9-) showed a sensitivity and a specificity of 100%. Therefore, the MARS model was applied to the modeling of the influence of several clinical and molecular variables on the prognosis of salivary gland cancers with success. A self-organizing map (SOM) is a type of neural network what was used here to determine a prognostic model composed for four variables: N, M, number of recurrences and tumor type. The sensitivity of this model was that of 97%, and its specificity was that of 94.7%.

Multi-GPU Development of a Neural Networks Based Reconstructor for Adaptive Optics

Aberrations introduced by the atmospheric turbulence in large telescopes are compensated using adaptive optics systems, where the use of deformable mirrors and multiple sensors relies on complex control systems. Recently, the development of larger scales of telescopes as the E-ELT or TMT has created a computational challenge due to the increasing complexity of the new adaptive optics systems. The Complex Atmospheric Reconstructor based on Machine Learning (CARMEN) is an algorithm based on artificial neural networks, designed to compensate the atmospheric turbulence. During recent years, the use of GPUs has been proved to be a great solution to speed up the learning process of neural networks, and different frameworks have been created to ease their development. The implementation of CARMEN in different Multi-GPU frameworks is presented in this paper, along with its development in a language originally developed for GPU, like CUDA. This implementation offers the best response for all the presented cases, although its advantage of using more than one GPU occurs only in large networks.

An Artificial Neural Network Model for the Prediction of Bruxism by Means of Occlusal Variables

The objective of the present work was to create an artificial neural network model able to classify individuals suffering from bruxism in clenching and grinding patients according to the value of certain occlusal variables and other parameters. Patients suspected of bruxism represent a very heterogeneous group. Some require immediate treatment while others, with only minor disorders, may not need treatment at all.

Comparison of the Periimplant Bone Stress Distribution on Three Fixed Partial Supported Prosthesis Designs Under Different Loading. A 3D Finite Element Analysis

Background: Despite high success and survival rates of implant supported prosthesis therapy, biomechanical complications such as periimplant bone resorption continue to exist due to occlusal overloading.

New adaptive optics Tomographic Pupil Image reconstructor based on convolutional neural networks

Astronomical images taken from large ground-based telescopes requires techniques as Adaptive Optics in order to improve their spatial resolution. In this work are presented computational results from a modified curvature sensor, the Tomographic Pupil Image Wavefront Sensor (TPI-WFS), which measures the turbulence of the atmosphere, expressed in terms of an expansion over Zernike polynomials. Convolutional Neural Networks (CNN) are presented as an alternative to the TPI-WFS reconstruction. This technique is a machine learning model of the family of Artificial Neural Networks (ANN), which are widely known for its performance as modeling and prediction technique in complex systems. Results obtained from the reconstruction of the networks are compared with the TPI-WFS reconstruction by estimating errors and optical measurements (Root Mean Square error, Mean Structural Similarity and Strehl ratio). In general, CNN trained as reconstructor showed slightly better performance than the conventional reconstruction in TPI-WFS for most of the turbulent profiles, but it made significant improvements for higher turbulent profiles that have the lowest r0 values.

EPIK - Virtual Rehabilitation Platform Devised to Increase Self-reliance of People with Limited Mobility

In this paper we describe a virtual rehabilitation platform designed to improve balance of people with physical impairment using the Microsoft® Kinect® sensor. Different types of users can interact with the platform: Administrators, therapists, and final users (patients), using their own interfaces and modules. Six modules have been designed: Profile, Administrator, Evaluation, Therapist, Game and Results; but only four have been implemented so far: Administrator, Evaluation, Therapist and Game. The Administrator’s module is used to generate a database of exercises. The Therapist’s module allows therapists to configure the game training session using combinations of exercises from the database. The patients’ or game module includes a 3D immersive environment, where they perform the prescribed rehabilitation exercises, previously configured by a therapist. The platform is in its first beta version and ready to be tested.