Fusion of Deep Convolutional Neural Networks

In recent years, the concept of big data has become a more prominent research topic as the volume of data and the rate at which it is produced are increasing exponentially. By 2020 the amount of data being stored is estimated to be 44 Zettabytes and currently over 31 Terabytes of data is being generated every second. Algorithms and applications must be able to effectively scale to the volume of data being generated. One such application that has excelled due to the surge in Big Data is the Convolutional Neural Network. The breakthroughs in the development of Graphical Processing Units have led to the advancements in the state-of-the-art on tasks such as image classification and speech recognition. These multi-layered convolutional neural networks are very large, complex and require significant computational resources to train and evaluate models. In this paper, we explore several novel architectures for the fusion of multiple convolutional neural networks, including stacked representation fusions and mixed model fusion. We differ from existing fusion methods in that our approaches take in the raw outputs of several CNN models and use classifiers as fusers. Other methods typically hand-craft the fusion or have used the original input space as the fusion method. Advancements in this area will better enable the leveraging of the vast amount of pretrained models and improve accuracy of these models. The approaches generated are application agnostic and will apply across a breadth of tasks.