VLSI implementation using fully connected neural networks for energy consumption over neurons. (August 2022)

• Record Type: Journal Article
• Authors: Mehbodniya, Abolfazl; Kumar, Ravi; Bedi, Pradeep; Nandan Mohanty, Sachi; Tripathi, Rohit; Geetha, A.
• Is Part Of: Sustainable energy technologies and assessments. Volume 52:Part A(2022)
• Publication Date: 2022-08
• Subjects: Neural Network -- Deep learning -- Energy consumption -- VLSI; Renewable energy sources -- Periodicals; Energy development -- Technological innovations -- Periodicals; Electric power production -- Periodicals; Energy storage -- Periodicals; 333.79
• DOI: 10.1016/j.seta.2022.102058
• Languages: English
• ISSNs: 2213-1388

Abstract

Highlights: The number of connections within fully connected systems up to 90% enhancing the quality of three popular data sets. To dramatically reduce DNN memory requirements. Validating the usefulness of the suggested networks as a regularization term to avoid unbalanced class systems. To employ restricted bandwidth off-chip memory and waste a great deal of sustainable energy. Abstract: Increasingly, the capacity of artificial neural networks to identify and express top-level concepts in sets of data has become quite significant. Deep learning models like fully connected layer machine learning have demonstrated outstanding results in range of identification and verification applications. It's technology deployments suffered as from complexity and huge silicone areas with energy consumption. Memory accesses, the bulk from which happens in fully connected systems, comprise the energy usage of machine learning. It has many of the depths of neural network. Throughout this work, researchers suggest densely connected networks by demonstrating that there is indeed a drop in the number of links inside fully connected systems of up to 90% enhancing the quality of three popular sets of data. To decrease the memory needs have suggested low-connected grid networks, researchers present perhaps an effective hardware implementation based on the linear response registers. In comparison to the traditional design of fully connected layer neural network models, the suggested design could save up to 90 percent of storage. Findings of execution further reveal that artificial neurons of suggested weak network connections have a power consumption decrease up to 84 percent compared to one single atom of fully connected layer machine learning.