SSGD: A Safe and Efficient Method of Gradient Descent

Table 1

The test accuracy of compared algorithms on MNIST.

	SGD ()	SGDm ()	Adam	SSGD (m)
16	0.9781 ()	0.9778 ()	0.9767	0.9832 (4)
32	0.9702 ()	0.9768 ()	0.9850	0.9876 (8)
64	0.9794 ()	0.9792 ()	0.9842	0.9900 (8)
128	0.9676 ()	0.9780 ()	0.9896	0.9901 (16)
256	0.9755 ()	0.9804 ()	0.9855	0.9877 (16)
512	0.9665 ()	0.9789 ()	0.9814	0.9861 (16)
1024	0.9738 ()	0.9749 ()	0.9778	0.9869 (16)
2048	0.9763 ()	0.9717 ()	0.9894	0.9886 (64)
4096	0.9703 ()	0.9806 ()	0.9788	0.9878 (64)
8192	0.9785 ()	0.8994 ()	0.9753	0.9855 (64)