Feature discovery and visualization of robot mission data using convolutional autoencoders and Bayesian nonparametric topic models

摘要

The gap between our ability to collect interesting data and our ability toanalyze these data is growing at an unprecedented rate. Recent algorithmicattempts to fill this gap have employed unsupervised tools to discoverstructure in data. Some of the most successful approaches have usedprobabilistic models to uncover latent thematic structure in discrete data.Despite the success of these models on textual data, they have not generalizedas well to image data, in part because of the spatial and temporal structurethat may exist in an image stream. We introduce a novel unsupervised machine learning framework thatincorporates the ability of convolutional autoencoders to discover featuresfrom images that directly encode spatial information, within a Bayesiannonparametric topic model that discovers meaningful latent patterns withindiscrete data. By using this hybrid framework, we overcome the fundamentaldependency of traditional topic models on rigidly hand-coded datarepresentations, while simultaneously encoding spatial dependency in our topicswithout adding model complexity. We apply this model to the motivatingapplication of high-level scene understanding and mission summarization forexploratory marine robots. Our experiments on a seafloor dataset collected by amarine robot show that the proposed hybrid framework outperforms currentstate-of-the-art approaches on the task of unsupervised seafloor terraincharacterization.