Nested Models and Nonparametric LSTMs in Vision-Based Autonomous Driving and Developing an R Package for Bayesian-Optimized Deep Learning

Abstract

Most current autonomous vehicle systems largely rely on Light Detection And Ranging (LIDAR), but there is much room for development in purely vision-based systems. Recently, a DeepDriving paradigm of learning a ordance indicators from input images was suggested. We extend this model by analyzing the temporal and sequential nature of the input images. To do so, we apply convolutional feature extraction and examine three problems: the usefulness of history, the lag selection problem, and the viability of complex models with history. We train recurrent, or "nested" statistical models of varying complexity, and our results demonstrate that the sequential aspect of this problem is highly important and appropriate for such recurrent statistical models.