Table 3 Absolute improvement (%) wrt the PCA baseline when bases are learned from the unlabeled GT data sets
From: High level feature extraction for the self-taught learning algorithm
Training | Dictionary size K | |||
|---|---|---|---|---|
Set | 100 | 200 | 300 | 500 |
GT-50 | ||||
IS-20 | 13.6 | 16.1 | 15.9 | 16.8 |
IS-50 | 11.1 | 11.3 | 13.5 | 14.0 |
IS-100 | −0.7 | 0.4 | 2.0 | 1.7 |
IS-250 | −0.6 | −0.3 | 0.0 | −0.6 |
GT-100 | ||||
IS-20 | 13.3 | 15.7 | 15.4 | 16.7 |
IS-50 | 10.5 | 10.9 | 13.7 | 13.7 |
IS-100 | −0.9 | 1.0 | 1.7 | 1.5 |
IS-250 | −0.4 | −0.5 | 0.0 | −0.3 |
GT-250 | ||||
IS-20 | 13.8 | 15.9 | 18.1 | 17.3 |
IS-50 | 10.5 | 11.4 | 15.7 | 13.5 |
IS-100 | −0.6 | 0.9 | 2.5 | 1.6 |
IS-250 | −0.9 | 0.1 | 0.1 | −0.5 |
GT-500 | ||||
IS-20 | 14.4 | 15.9 | 15.2 | 16.8 |
IS-50 | 11.1 | 11.2 | 14.0 | 14.0 |
IS-100 | −0.4 | 1.4 | 2.1 | 1.5 |
IS-250 | −0.9 | 0.1 | 0.2 | −0.3 |