A Hybrid Deep Neural Network System And Its Applications

Abstract

This study presents the design and implementation of need based hybrid deep neural networks. It shows a synergistic integration with other approaches which enables in-depth modeling of deep networks for various applications. Based on the nature and density of information provided at the input, variable size receptive fields help in achieving high performance. This has been demonstrated on breast cancer histopathological (BreaKHis) dataset. Inspired by this, a heterogeneous and modular approach based hybrid network (HMDNN) has been designed to cover degree of information embedded in a given input. With this approach, the problem of diabetic retinopathy has been solved effectively by showing promising results. newline Next, an integration of compact evolutionary algorithm in a cohesive manner allows us to design cost-effective optimization approach for deep networks. The proposed hybrid evolutionary gradient descent (HyEGD) approach offers simultaneous learning of structure and weight parameters for the class of convolutional neural networks. Experiments have been performed to understand the effect of parameters which highly influence the performance as well as structure of convolutional networks. The proposed approach has been validated on various benchmark datasets. A careful integration of the architectural economy parameters in the cost functions resulted in the flexibility in designing of economical structures in terms of size as well as performance. newlineThe present work also introduces a novel approach to design a simpler topology for DenseNet by introducing sparse connections in in the structure thereby resulting in sparse blocks. The simpler design of the network teamed up with the hybrid evolutionary gradient descent algorithm helped in reducing the network parameters although by maintaining the network performance. In this case, the following benchmark datasets have been used to show the effectiveness of the proposed approach; Malaria Parasite Detection, Human Epithelial Type 2 dataset, Epilepsy dataset, Colorectal Cancer dataset, Retinal Optical Coherence Tomography dataset, Primary microRNA dataset and COVID-19 Genome Sequence dataset. newline newline