Table 2.

Linear evaluation results with labeled STL10, using a base encoder trained on unlabeled STL10.

Loss	Learning	Update	Acc.
Contr. Hinge	E2E	BP	70.13%
		RF	55.01%
	GLL	BP/URF	68.26% / 68.80%
		RF	60.00% / 60.50%
	RLL	BP/URF	64.61%
		RF	55.14%
CLAPP	E2E	BP	71.88%
CLAPP	GLL	N/A	68.74%
Rnd. encoder	46.78%

The contrastive hinge loss (Contr. Hinge) with 5 negatives is compared with patch-based CLAPP in both end-to-end (E2E) and greedy layer-wise learning (GLL). We also compare the results from randomized layer-wise learning (RLL), and four updating methods, backpropagation (BP), updated random feedback (URF), and random feedback (RF) when using contrastive hinge loss. Underlined results were obtained with 1 negative.