A Model for Reduction of Time and Space Complexity on Edge Devices

In the realm of edge computing, a paradigm emphasizing decentralized computational tasks, the interplay between time and space complexity holds immense significance. Time complexity denotes the duration required for an algorithm's execution, while space complexity concerns the memory or storage demand throughout the process. The evaluation entails a comparative analysis between a conventional non-quantized model and its quantized counterpart, focusing on accuracy, memory utilization, and runtime. The non-quantized model exhibits commendable learning performance, achieving a 96% accuracy rate during training but experiencing a marginal decrease to 90% in testing. Conversely, the quantized model sustains competitive accuracy, attaining 98% in both training and testing phases. The architecture of the quantized model, characterized by diminished numerical precision, emerges as a pivotal factor in minimizing both memory footprint and computational requirements. Graphical analyses unveil that despite a slight increase in loss during validation, the quantized model displays robust learning and generalization capabilities from the training dataset. The comparative analysis emphasizes the benefits of quantization, emphasizing decreased memory utilization (3kb), faster runtime, and, in specific cases, improved accuracy (96%). This thesis provides valuable perspectives on the effectiveness of quantization in optimizing Convolutional Neural Network (CNN) models for deployment on edge devices with limited resources. The evaluation metrics employed include memory usage reduction, runtime speed, and accuracy enhancement 96%.