Topics in Deep Learning Classification

Koo, Jaehoon

doi:https://doi.org/10.21985/n2-6bcb-5r93

Work

Topics in Deep Learning Classification

Public

Download PDF

Download All Files (.zip)

In machine learning, classification that assigns a label to a sample is a fundamental problem and serves a building block for various applications of artificial intelligence such as speech recognition, sentimental analysis, and image recognition. During the last years, deep learning rejuvenates artificial intelligence; in particular, it leads to tremendous progress in classification tasks. In this study, we develop enhanced deep learning methodologies for supervised classification. We also explore training schemes and implementations of the models using high-end computing machines. Furthermore, we study an interesting variant of the classification problem, called inverse classification that explores interpretability of classification models. This dissertation consists of three chapters, 1) Improved Classification Methods Based on Deep Belief Networks (DBN), 2) Combined Convolutional and Recurrent Neural Networks for Hierarchical Classification of Images, and 3) A New Framework for Inverse Classification Using Mixed Integer Programming. In the first chapter, we explore how to incorporate unsupervised learning methods in supervised classification. Generative models are commonly used to initialize classifiers before fine-tuning. Typically, this requires solving separate unsupervised and supervised learning problems. In this work, we focus on DBN, which is a widely used unsupervised model. We develop several supervised models incorporating DBN in order to improve the two-phase learning strategy. The improvements over two-phase are consistent. In the second chapter, we focus on hierarchical classification of images. Object classes have known hierarchical relations, and classifiers exploiting these relations can perform better. To incorporate this perspective, we develop a combined model for classification that extracts hierarchical representations of images by a convolutional neural network and learns a tree of label paths to predict a final label of images by a recurrent neural network. The proposed model leads to image classification that captures the hierarchical characteristics of the classes. In the third chapter, we shift our attention to studying interpretability of classification models rather than improving classification accuracy. We study an inverse classification problem that is a machine learning task designed to identify small changes needed in input features of an instance to adjust its associated prediction as desired. To solve this problem, we formulate a constrained mixed integer programming problem and design an associated algorithm based on Lagrangian and subgradient methods.

Creator