Developing a Distributed Computing Framework for Accelerating Deep learning for Google Street View

Digital maps has been increasingly used in the past few years, and street views in maps potentially draw people attention before they arrive their destination. Accordingly, study the human opinion to a street view of specific address is attractive as it is able to provide useful information for sociology and psychology in the near future. However, the original deep learning (DL) model used in processing street views barely can release all GPUs computational power, and lack of substantial scalability and stability when encountering relatively large dataset. Therefore, we build a distributed framework to accelerate the DL model in server clusters which improves the GPU utility significantly. Along with the framework, an application is developed to fetch historical street views which will be fed into the DL model from Google API. A data analyzing toolkit is also implemented to enhance the DL model results collecting and summarizing. With this developed framework, the processing efficiency in server clusters of the DL model is noticeably improved which brings considerably shorter prediction time.

Major Advisor: Lizhong Chen
Committee: Yue Zhang
Committee: Perry Hystad