Image Miner : an architecture to support deep mining of images

Abstract

In this thesis, I designed a cloud based system, called ImageMiner, to tune parameters of feature extraction process in a machine learning pipeline for images. Feature extraction is a key component of the machine learning pipeline, and tune its parameters to extract the best features can have significant effect on the accuracy achieved by the machine learning system. To enable scalable parameter tuning, I designed a master-slave architecture to run on the Amazon cloud. To overcome the computational bottlenecks due to large datasets, I used a data parallel approach where each worker runs independently on a subset of data. The worker uses a Gaussian Copula Process to tune parameters and determines the best set of parameters and model to use.