Robots, Autonomous Vehicles and Advanced Computer Applications

Deep learning is the science of training large artificial neural networks. Deep neural networks (DNNs) can have hundreds of millions of parameters allowing them to model complex functions such as nonlinear dynamics. They form compact representations of state from raw, high-dimensional, multimodal sensor data commonly found in robotic systems and unlike many machine learning methods, they do not require a human expert to hand-engineer feature vectors from sensor data at design time. DNNs can, however, present particular challenges in physical robotic systems, where generating training data is generally expensive, and sub-optimal performance in training poses a danger in some applications. Yet, despite such challenges, robotcists are finding creative alternatives, such as leveraging training data via digital manipulation, automating training, and employing multiple DNNs to improve performance and reduce training time.