A Separation Principle for Control in the Age of Deep Learning

We review the problem of defining and inferring a state for a control system based on complex, high-dimensional, highly uncertain measurement streams, such as videos. Such a state, or representation, should contain all and only the information needed for control and discount nuisance variability in the data. It should also have finite complexity, ideally modulated depending on available resources. This representation is what we want to store in memory in lieu of the data, as it separates the control task from the measurement process. For the trivial case with no dynamics, a representation can be inferred by minimizing the information bottleneck Lagrangian in a function class realized by deep neural networks. The resulting representation has much higher dimension than the data (already in the millions) but is smaller in the sense of information content, retaining only what is needed for the task. This process also yields representations that are invariant to nuisance factors and have maximally independent components. We extend these ideas to the dynamic case, where the representation is the posterior density of the task variable given the measurements up to the current time, which is in general much simpler than the prediction density maintained by the classical Bayesian filter. Again, this can be finitely parameterized using a deep neural network, and some applications are already beginning to emerge. No explicit assumption of Markovianity is needed; instead, complexity trades off approximation of an optimal representation, including the degree of Markovianity.

Probability of Heat Intolerance: Standardized Interpretation of Heat-Tolerance Testing Results Versus Specialist Judgment

Context: The heat-tolerance test (HTT) is a screening tool for secondary prevention of exertional heat illness by the Israel Defense Forces. To discern participant tolerance, recruits are exposed to intermediate environmental and exercise stresses, and their physiological responses, core temperature, and heart rate are monitored. When their physiological measures rise at a higher rate or exceed the upper levels of absolute values compared with other participants, heat intolerance (HI) is diagnosed. Objective: To develop a mathematical model to interpret HTT results and provide a quantitative estimate of the probability of heat tolerance (PHT). Design: Cross-sectional study. Setting: Warrior Health Research Institute. Patients or Other Participants: The HTT results of 175 random individuals tested after an episode of exertional heat illness were classified qualitatively and then divided into training (n = 112) and testing (n = 63) datasets. All individuals were male soldiers (age range = 18–22 years) who had sustained an episode of definitive or suspected exertional heat stroke. Main Outcome Measure(s): Based on the decision algorithm used by the Israel Defense Forces for manual interpretation of the HTT, we designed a logistic regression model to predict the heat-tolerance state. The model used a time series of physiological measures (core temperature and heart rate) of individuals to predict the manually assigned diagnosis of HT or HI. It was initially fitted and then tested on 2 separate, random datasets. The model produced a single value, the PHT, and its predictive ability was demonstrated by prediction-density plots, receiver operating characteristic curve, contingency tables, and conventional screening test evaluation measures. Results: According to prediction-density plots of the testing set, all HT patients had a PHT of 0.7 to 1. The receiver operating characteristic curve plot showed that PHT was an excellent predictor of the manual HT interpretations (area under the curve = 0.973). Using a cutoff probability of 0.5 for the diagnosis of HI, we found that PHT had sensitivity, specificity, and accuracy of 100%, 90%, and 92.06%, respectively. Conclusions: The PHT has the potential to be substituted for manual interpretation of the HTT and to serve in a variety of clinical and research applications.

A Comparison of SLAM Prediction Densities Using the Kolmogorov Smirnov Statistic

Accurate pose and trajectory estimates, are necessary components of autonomous robot navigation system. A wide variety of Simultaneous Localization and Mapping (SLAM) and localization algorithms have been developed by the robotics community to cater to this requirement. Some of the sensor fusion algorithms employed by SLAM and localization algorithms include the particle filter, Gaussian Particle Filter, the Extended Kalman Filter, the Unscented Kalman Filter, and the Central Difference Kalman Filter. To guarantee a rapid convergence of the state estimate to the ground truth, the prediction density of the sensor fusion algorithm must be as close to the true vehicle prediction density as possible. This paper presents a Kolmogorov–Smirnov statistic-based method to compare the prediction densities of the algorithms listed above. The algorithms are compared using simulations of noisy inputs provided to an autonomous robotic vehicle, and the obtained results are analyzed. The results are then validated using data obtained from a robot moving in controlled trajectories similar to the simulations.

OTIMIZAÇÃO DE CALIBRAÇÕES BASEADAS EM ESPECTROSCOPIA NO INFRAVERMELHO PRÓXIMO PARA ESTIMATIVA DE PROPRIEDADES DA MADEIRA DE Eucalyptus

A espectroscopia no infravermelho próximo (NIRS) é uma técnica não-destrutiva, rápida e utilizada para avaliação, caracterização e classificação de materiais, sobretudo de origem biológica. A obtenção de informações contida nos espectros NIR é complexa e requer a utilização de métodos quimiométricos. Assim, por meio de regressão multivariada, os espectros de absorbância podem ser associados às propriedades da madeira, tornando possível a sua predição em amostras desconhecidas. Existem algumas ferramentas quimiométricas que melhoram o ajuste dos modelos preditivos. Assim, o objetivo deste trabalho foi simular regressões dos mínimos quadrados parciais baseados nas informações espectrais e de laboratório e estudar a influência da aplicação de tratamentos matemáticos, do descarte de amostras anômalas e da seleção de comprimentos de onda no ajuste dos modelos para estimativa da densidade básica e do módulo de elasticidade em ensaio de compressão paralela às fibras da madeira de Eucalyptus. A aplicação da primeira e segunda derivada nos espectros, o descarte de amostras anômalas e a seleção de algumas das variáveis espectrais melhorou significativamente o ajuste do modelo, reduzindo o erro padrão e aumentando o coeficiente de determinação e a relação de desempenho do desvio.Palavras-chave: Espectroscopia no infravermelho próximo; predição; densidade básica; MOE; madeira; Eucalyptus. AbstractOptimization of calibrations based on near infrared spectroscopy for estimation of Eucalyptus wood properties. Near infrared spectroscopy (NIRS) is a non-destructive technique used for rapid evaluation, characterization and classification of biological materials. The extraction of the information contained in the NIR spectrum is complex and requires the use of chemo metric methods. Thus, by means of multivariate regression, the absorbance spectra are correlated to wood properties, making possible the prediction in unknown samples. There are some chemo metric tools that can improve the adjustment of the predictive models. The aim of this work was to simulate partial least squares regression based on NIR spectra and laboratory data and to study the influence of the application of mathematical treatment, the removal of outliers and the wavelengths selection in the adjustment of models to estimate the density and modulus of elasticity in Eucalyptus wood. The use of the first and second derivative spectra, the disposal of outliers, and the variables selection improved significantly the model fit, reducing the standard error and increasing the coefficient of determination and the ratio of performance to deviation.Keywords: Near infrared; spectroscopy; prediction; density; MOE; wood; Eucalyptus.