MCDD: Multi-class Distribution Model for Large Scale Classification

Mayanka Chandrashekar, Yugyung Lee

Abstract: A parallel and distributed machine learning framework are in need to deal with a large amount of data. We have seen unsatisfactory classification performance especially with increasing the number of classes. In this paper, we propose a distributed deep learning framework, called Multi-Class Discriminative Distribution (MCDD) that aims to distribute classes while improving the accuracy performance of the deep learning models with large scale datasets. The MCDD framework works on an evidence-based learning model for the optimal distribution of classes by computing a misclassification cost (i.e., confusion factor). These observations about learning attempts have been used to extend a classifier into a classification model hierarchy by learning an optimal distribution of classes. As a result, a distributed deep neural network model with multi-class classifiers (MCDD) was built to optimize the accuracy and performance of the learning process. The MCDD model runs on parallel environments, such as Apache Spark and Tensor Flow using large real-world datasets (Caltech-101, CIFAR-100, ImageNet-1K) showing that MCDD can build a class distribution model with higher accuracy compared to existing models.