Computer Sciences and Information Technology
Title : | Theoretical Evaluation of Different Strategies (Choice of Kernel, Striding and Pooling) During Implementing Deep Convolution Neural Network (DCNN) Pipeline |
Area of research : | Computer Sciences and Information Technology |
Principal Investigator : | Dr. Shovan Barma, Indian Institute Of Information Technology, Assam |
Timeline Start Year : | 2023 |
Timeline End Year : | 2026 |
Contact info : | shovan.barma@yahoo.com |
Details
Executive Summary : | Deep learning (DL) has shown significant performance in various tasks, including classification, object detection, and segmentation, in various domains such as computer vision, natural language processing, and medical image processing. The foundation of DL, deep convolution neural network (DCNN) pipelines, is structured in hierarchical layers that perform linear transformations through convolution operations and nonlinear activation functions. However, there is limited theoretical development of DCNN compared to architectural modifications for domain-specific applications. The fundamental DCNN architecture includes several convolution layers, pooling layers, and fully connected dense layers. The pragmatic success of any DL lies in its future extraction mechanism through convolution operation and the number of concatenated layers. Currently, these operations are performed empirically without proper mathematical reasoning. This work aims to explore the mathematical reasons behind the optimum abstract level features generated during implementing DCNN pipelines. The mathematical foundation will be established based on a statistical learning framework, and two popular image datasets for deep leaning applications, CIFAR and MNIST, will be used for proper validation. The DCNN will be implemented on a computer platform using Python, and the project will be a small step towards developing the mathematical foundation of deep leaning. |
Total Budget (INR): | 6,60,000 |
Organizations involved