en-es
en
0.25
0.5
0.75
1.25
1.5
1.75
2
Learning with Marginalized Corrupted Features
Published on Aug 26, 20134724 Views
We propose a new framework for regularization, called marginalized corrupted features, that reduces over fitting by increasing the robustness of the model to data corruptions.
Related categories
Chapter list
Learning with Marginalized Corrupted Features00:00
Classification of text and image data00:19
Empirical risk minimization00:43
Regularization by priors - 101:47
Regularization by priors - 202:55
Regularization by priors - 304:38
Movie reviews04:51
Regularization by corruption - 105:24
Regularization by corruption - 206:14
Regularization by corruption - 307:03
Simple approach - 108:48
Simple approach - 209:40
Marginalized Corrupted Features09:42
Quadratic loss - 110:20
Quadratic loss - 211:13
Exponential loss12:16
Moment-generating functions13:17
Blankout: Ensemble interpretation - 113:27
Blankout: Ensemble interpretation - 213:48
Blankout: Ensemble interpretation - 313:59
Blankout: Ensemble interpretation - 414:05
Logistic loss14:59
Using MCF in practice15:52
Experimental setup16:15
Experiment 1: Document classification - 117:47
Experiment 1: Document classification - 217:49
Experiment 1: Document classification - 319:19
Experiment 2: Image classification20:13
Experiment 3: “Nightmare at test time” - 120:56
Experiment 3: “Nightmare at test time” - 221:56
Conclusions23:09
Thank you! Questions?23:57