[Retracted] Gradient Descent Optimization in Deep Learning Model Training Based on Multistage and Method Combination Strategy

Table 8

Performance of the proposed method, SGD.

	ResNet-20 on Cafri-10		LSTM on IMDB
	Val-loss	Val-acc	Val-loss	Val-acc
SGD + SGD	1.0178	0.6948	0.6919	0.5570
SGD + (SGD + M)	1.0763	0.7134	0.4408	0.7971
SGD + (SGD + d)	0.9607	0.7168	0.6890	0.5777
SGD + (SGD + M + d)	0.9040	0.7557	0.4353	0.7982
SGD + RMSprop	0.9408	0.7419	0.4287	0.8367
SGD + (RMSprop + d)	1.0131	0.7298	0.4342	0.8237
SGD + Adam	0.8751	0.7641	0.9210	0.8100
SGD + (Adam + d)	1.0692	0.7274	0.8172	0.8130

m: Momentum, D: decay by 1e − 6 every iteration, and “()”: take methods at the same timepiece. The bold values represent the best results.