DenseNet: Assessing Limitations in Connectivity and Memory Efficiency
Mohammad Hamza Parvez Alam Siddiqui · INTERANTIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING AND MANAGEMENT · 2024
Abstract—Despite the prevailing trend of introducing skip connections in deep neural networks for enhanced depth and training efficiency, this paper critically examines the limitations of Dense Convolutional Network (DenseNet), a prominent state- of-the-art architecture. The analysis highlights the drawbacks of connecting each layer to every other layer in a feed-forward manner. Additionally, this paper systematically explores the shortcomings and challenges associated with DenseNet, shedding light on its drawbacks. The study culminates in the identification of critical areas for improvement and suggests avenues for optimizing memory utilization during training. Ultimately, this paper aims to provide a comprehensive understanding of the inherent limitations in DenseNet architecture and offers insights into potential advancements in deep neural network design.