Information Transfer Strategies for Performant Learning Tasks in Deep Learning

The advancements in deep learning have revolutionized the landscape of various applications, yet deep neural networks face challenges such as high computational complexity and sensitivity to noisy input. This study presents efficient strategies for learning tasks based on information transfer to mitigate these issues. The first part of the project introduces a novel network compression technique Information Consistent Pruning (InCoP), an Iterative Magnitude Pruning (IMP) algorithm that utilizes flow-based metrics to optimize the pruning process without relying on accuracy for stopping criteria. InCoP significantly reduces the number of training epochs required, enhancing the efficiency of deep networks. In the subsequent phase, the study delves into the utilization of deep networks for object detection in adverse weather conditions, where input images are frequently distorted, presenting challenges for detection algorithms. Employing a knowledge distillation framework within a teacher-student architecture, the FogGuard facilitates the transfer of knowledge from the teacher network under clear weather conditions to the student network, enhancing object detection accuracy. Furthermore, leveraging perceptual loss demonstrates robustness against pixel-wise image alterations caused by adverse weather. Lastly, the exploration extends to utilizing Vision Transformers (ViT), focusing on Real-Time Detection Transformers (RT-DETR) for enhanced adaptability to adverse weather conditions. A novel design termed FogAware Attention enables the detection of weather effects on image patches and dynamically adjusts attention levels, showcasing potential advancements in weather-adaptive image analysis using ViT.