Deep learning ensembles for image understanding

Official URL: https://risc01.sabanciuniv.edu/record=b2579336_(Table of contents)

Deep neural networks have enhanced the performance of decision making systems in many applications, including image understanding. Further performance gains can be achieved by using ensemble methods, which are shown to be powerful tools for various classification and regression tasks. This dissertation consists of two parts. The first part is devoted to studying the face attributes classification problem. We introduce several novel approaches for this problem, achieving state-of-art results on CelebA and LFWA datasets: i) we use the multi-task learning (MTL) framework for multiple attributes classification for scalability, where base learners are grouped according to the location of the attribute on the face and share weights. Giving information about the location of an attribute as prior information is shown to speed up the learning process and lead to increased accuracy. ii) we introduce a novel ensemble learning technique within the deep learning model itself (within-network ensemble), showing increased performance at almost the same time complexity of a single model. iii) we propose a new framework called Deep-RankSVM for relative attribute classification (comparing the attribution expression on two photographs) adapting the SVM formulation to deep rank learning. The second part is devoted to analyzing the suitability of different state-of-art design strategies for constructing ensembles of deep networks. We propose the Error Correcting Output Codes (ECOC) framework as a novel deep learning ensemble method, and show that it can be used with the MTL framework for arbitrary accuracycomplexity trade-off. We carry out an extensive comparative study between the introduced ECOC designs and the state-of-the-art ensemble techniques such as ensemble averaging and gradient boosting decision trees, on several datasets. In the rest of the dissertation, we discuss general applications of the proposed ensemble techniques that include skin lesion classification and plant identification.