Managing Deep Learning Uncertainty for Unmanned Systems

Abstract

Unmanned or Autonomous Systems act in an uncertain world. In order to plan and make decisions, autonomous systems can only rely on noisy perceptions and approximated models. The data collected from sensors is inherently uncertain due to noise, incompleteness, and inconsistency. The analysis of such massive amounts of data requires advanced analytical techniques for efficiently reviewing and/or predicting future courses of action with high precision and advanced decision-making strategies. Due to the amount, variety, and speed of data increases, so too does the uncertainty inherent within, leading to a lack of confidence in the resulting analytics process and decisions made thereof. Under these conditions, deep learning algorithms can be fully integrated into robotic systems only if a measure of prediction uncertainty is available. Prediction uncertainty in deep neural networks generally derives from two sources: data uncertainty and model uncertainty. The former arises because of noise in the data, usually caused by the sensors’ imperfections. The latter instead is generated from unbalances in the training data distribution. The article describes the developing fully autonomous vehicles at three levels, namely, perception, planning and control. Ideas from artificial intelligence have also been used to solve planning problems for autonomous vehicles. Autonomous vehicles have to deal with dynamic, non-stationary- and highly unpredictable operational environments. For the right understanding of this process, is described the best-known control strategies, map representation techniques, interaction with external systems, including pedestrian interactions under an informational conception of Internet of Things and Big Data Uncertainty and Uncertainty in Machine Learning, The article shows an analysis of the most promissory methods for treating the deep leaning uncertainty: Bayesian Deep Learning, which is an intersection between deep learning and Bayesian probability approaches and the Fuzzy logic machine learning, in an Internet of Things framework. For the validation of theoretical explanations, it was proposed some simulations tools for visualizing and interpreting the uncertainty in the previously mentioned models and in the internet of things framework in different environments, useful for teaching and research.