Skip to main content
. 2022 Mar 21;16:789253. doi: 10.3389/fncom.2022.789253

Table 4.

Test accuracy on transformed MNIST using the pre-trained networks with SSL on EMNIST letters.

Loss Learning Update Acc. on transformed digits
Contr. Hinge E2E BP 80.20%
DTP 81.37%
RF 75.89%
GLL BP 81.31%
RF 80.86%
RLL BP 80.31%
RF 78.52%
SimCLR E2E BP 81.96%
Supervised 75.6%
Rnd. encoder 62.66%

Linear classifiers/supervised networks trained on original MNIST.