Deep learning feature extraction with limited data and processing

Abstract

In this project, we will explore the use of state-of-the-art methods to speed-up training and testing times of deep learning models. We are also interested in developing novel techniques for training these models with a limited amount of data. For speed-up, we will explore custom hardware such as Graphical Processing Units (GPUs) for training and embedded hardware for testing. We will also explore ways to enhance an imaging database of unscreened microorganisms that we will gather in this project to make it more effective for training models. Deep learning models are known to be data hungry, and we will explore image transformations as well as transfer learning techniques to introduce diversity and reduce the dependency of data. We will assess the efficiency of the proposed techniques as the increase of accuracy and reduction of training and test times compared to the deep learning models trained directly from the image database. (AU)