Matches in SemOpenAlex for { <https://semopenalex.org/work/W2894770203> ?p ?o ?g. }
- W2894770203 abstract "Deep networks often perform well on the data distribution on which they aretrained, yet give incorrect (and often very confident) answers when evaluatedon points from off of the training distribution. This is exemplified by theadversarial examples phenomenon but can also be seen in terms of modelgeneralization and domain shift. Ideally, a model would assign lower confidenceto points unlike those from the training distribution. We propose a regularizerwhich addresses this issue by training with interpolated hidden states andencouraging the classifier to be less confident at these points. Because thehidden states are learned, this has an important effect of encouraging thehidden states for a class to be concentrated in such a way so thatinterpolations within the same class or between two different classes do notintersect with the real data points from other classes. This has a majoradvantage in that it avoids the underfitting which can result frominterpolating in the input space. We prove that the exact condition for thisproblem of underfitting to be avoided by Manifold Mixup is that thedimensionality of the hidden states exceeds the number of classes, which isoften the case in practice. Additionally, this concentration can be seen asmaking the features in earlier layers more discriminative. We show that despiterequiring no significant additional computation, Manifold Mixup achieves largeimprovements over strong baselines in supervised learning, robustness tosingle-step adversarial attacks, semi-supervised learning, and NegativeLog-Likelihood on held out samples." @default.
- W2894770203 created "2018-10-12" @default.
- W2894770203 creator A5018408976 @default.
- W2894770203 creator A5030891292 @default.
- W2894770203 creator A5031092503 @default.
- W2894770203 creator A5031185465 @default.
- W2894770203 creator A5083701330 @default.
- W2894770203 creator A5086198262 @default.
- W2894770203 creator A5088076429 @default.
- W2894770203 date "2018-09-27" @default.
- W2894770203 modified "2023-09-27" @default.
- W2894770203 title "Manifold Mixup: Learning Better Representations by Interpolating Hidden States" @default.
- W2894770203 cites W1496559305 @default.
- W2894770203 cites W2557579533 @default.
- W2894770203 cites W2592691248 @default.
- W2894770203 cites W2765407302 @default.
- W2894770203 cites W2788475065 @default.
- W2894770203 cites W2949117887 @default.
- W2894770203 cites W2950577311 @default.
- W2894770203 cites W2962879692 @default.
- W2894770203 cites W2963373786 @default.
- W2894770203 cites W2963836885 @default.
- W2894770203 cites W2964159205 @default.
- W2894770203 cites W2964253222 @default.
- W2894770203 hasPublicationYear "2018" @default.
- W2894770203 type Work @default.
- W2894770203 sameAs 2894770203 @default.
- W2894770203 citedByCount "17" @default.
- W2894770203 countsByYear W28947702032019 @default.
- W2894770203 countsByYear W28947702032020 @default.
- W2894770203 countsByYear W28947702032021 @default.
- W2894770203 crossrefType "posted-content" @default.
- W2894770203 hasAuthorship W2894770203A5018408976 @default.
- W2894770203 hasAuthorship W2894770203A5030891292 @default.
- W2894770203 hasAuthorship W2894770203A5031092503 @default.
- W2894770203 hasAuthorship W2894770203A5031185465 @default.
- W2894770203 hasAuthorship W2894770203A5083701330 @default.
- W2894770203 hasAuthorship W2894770203A5086198262 @default.
- W2894770203 hasAuthorship W2894770203A5088076429 @default.
- W2894770203 hasConcept C104317684 @default.
- W2894770203 hasConcept C111030470 @default.
- W2894770203 hasConcept C119857082 @default.
- W2894770203 hasConcept C127413603 @default.
- W2894770203 hasConcept C134306372 @default.
- W2894770203 hasConcept C153180895 @default.
- W2894770203 hasConcept C154945302 @default.
- W2894770203 hasConcept C177148314 @default.
- W2894770203 hasConcept C185592680 @default.
- W2894770203 hasConcept C21080849 @default.
- W2894770203 hasConcept C2777212361 @default.
- W2894770203 hasConcept C33923547 @default.
- W2894770203 hasConcept C37736160 @default.
- W2894770203 hasConcept C41008148 @default.
- W2894770203 hasConcept C529865628 @default.
- W2894770203 hasConcept C55493867 @default.
- W2894770203 hasConcept C63479239 @default.
- W2894770203 hasConcept C78519656 @default.
- W2894770203 hasConcept C95623464 @default.
- W2894770203 hasConcept C97931131 @default.
- W2894770203 hasConceptScore W2894770203C104317684 @default.
- W2894770203 hasConceptScore W2894770203C111030470 @default.
- W2894770203 hasConceptScore W2894770203C119857082 @default.
- W2894770203 hasConceptScore W2894770203C127413603 @default.
- W2894770203 hasConceptScore W2894770203C134306372 @default.
- W2894770203 hasConceptScore W2894770203C153180895 @default.
- W2894770203 hasConceptScore W2894770203C154945302 @default.
- W2894770203 hasConceptScore W2894770203C177148314 @default.
- W2894770203 hasConceptScore W2894770203C185592680 @default.
- W2894770203 hasConceptScore W2894770203C21080849 @default.
- W2894770203 hasConceptScore W2894770203C2777212361 @default.
- W2894770203 hasConceptScore W2894770203C33923547 @default.
- W2894770203 hasConceptScore W2894770203C37736160 @default.
- W2894770203 hasConceptScore W2894770203C41008148 @default.
- W2894770203 hasConceptScore W2894770203C529865628 @default.
- W2894770203 hasConceptScore W2894770203C55493867 @default.
- W2894770203 hasConceptScore W2894770203C63479239 @default.
- W2894770203 hasConceptScore W2894770203C78519656 @default.
- W2894770203 hasConceptScore W2894770203C95623464 @default.
- W2894770203 hasConceptScore W2894770203C97931131 @default.
- W2894770203 hasLocation W28947702031 @default.
- W2894770203 hasOpenAccess W2894770203 @default.
- W2894770203 hasPrimaryLocation W28947702031 @default.
- W2894770203 hasRelatedWork W1731081199 @default.
- W2894770203 hasRelatedWork W2120118379 @default.
- W2894770203 hasRelatedWork W2194775991 @default.
- W2894770203 hasRelatedWork W2765407302 @default.
- W2894770203 hasRelatedWork W2768915615 @default.
- W2894770203 hasRelatedWork W2891021639 @default.
- W2894770203 hasRelatedWork W2901453733 @default.
- W2894770203 hasRelatedWork W2902200315 @default.
- W2894770203 hasRelatedWork W2913881544 @default.
- W2894770203 hasRelatedWork W2949117887 @default.
- W2894770203 hasRelatedWork W2949277570 @default.
- W2894770203 hasRelatedWork W2963207607 @default.
- W2894770203 hasRelatedWork W2964910440 @default.
- W2894770203 hasRelatedWork W3026007863 @default.
- W2894770203 hasRelatedWork W3105463048 @default.
- W2894770203 hasRelatedWork W3118608800 @default.
- W2894770203 hasRelatedWork W3127699700 @default.
- W2894770203 hasRelatedWork W3132630012 @default.