In this section, the experimental details and implementation of a proposed method for optimizing deep neural networks (DNNs) are described. The method is evaluated on three different backbone networks and compared with two optimization methods, SGD and DEMON, using the CIFAR-100 dataset. The experiments are conducted on a NVIDIA GeForce RTX 3060 Laptop GPU using Python 3.10.9 and Pytorch 2.0.1.
The results show that the proposed method, referred to as PLP, and DEMON outperform SGD in terms of accuracy during the training process. The PLP method shows better optimization performance compared to DEMON as the training progresses. The method also demonstrates the ability to improve DNN training efficiency and performance compared to baseline models.
The sensitivity evaluation of the PLP method with different learning rates and batch sizes shows that the method performs better with smaller learning rates and exhibits good training performance with different batch size settings on various backbones. Overall, the PLP method shows good performance and low sensitivity to hyperparameter changes, verifying its effectiveness in optimizing DNN training performance and efficiency.
In conclusion, the experimental results suggest that the proposed PLP method is superior to SGD and comparable to state-of-the-art non-adaptive methods like DEMON in terms of convergence speed and accuracy. The method shows promise in enhancing DNN training efficiency and performance, with relatively low sensitivity to hyperparameter changes.
Source link
Source link: https://www.nature.com/articles/s41598-024-65691-0