Matches in SemOpenAlex for { <https://semopenalex.org/work/W2920293892> ?p ?o ?g. }
- W2920293892 abstract "It was empirically confirmed by Keskar et al.cite{SharpMinima} that flatter minima generalize better. However, for the popular ReLU network, sharp minimum can also generalize well cite{SharpMinimacan}. The conclusion demonstrates that the existing definitions of flatness fail to account for the complex geometry of ReLU neural networks because they can't cover the Positively Scale-Invariant (PSI) property of ReLU network. In this paper, we formalize the PSI causes problem of existing definitions of flatness and propose a new description of flatness - emph{PSI-flatness}. PSI-flatness is defined on the values of basis paths cite{GSGD} instead of weights. Values of basis paths have been shown to be the PSI-variables and can sufficiently represent the ReLU neural networks which ensure the PSI property of PSI-flatness. Then we study the relation between PSI-flatness and generalization theoretically and empirically. First, we formulate a generalization bound based on PSI-flatness which shows generalization error decreasing with the ratio between the largest basis path value and the smallest basis path value. That is to say, the minimum with balanced values of basis paths will more likely to be flatter and generalize better. Finally. we visualize the PSI-flatness of loss surface around two learned models which indicates the minimum with smaller PSI-flatness can indeed generalize better." @default.
- W2920293892 created "2019-03-11" @default.
- W2920293892 creator A5044802273 @default.
- W2920293892 creator A5070990160 @default.
- W2920293892 creator A5078600050 @default.
- W2920293892 creator A5080304958 @default.
- W2920293892 creator A5087695030 @default.
- W2920293892 date "2019-03-06" @default.
- W2920293892 modified "2023-09-27" @default.
- W2920293892 title "Positively Scale-Invariant Flatness of ReLU Neural Networks" @default.
- W2920293892 cites W1686810756 @default.
- W2920293892 cites W1811750039 @default.
- W2920293892 cites W2014384147 @default.
- W2920293892 cites W2029029543 @default.
- W2920293892 cites W2112796928 @default.
- W2920293892 cites W2194775991 @default.
- W2920293892 cites W2613904329 @default.
- W2920293892 cites W2626778328 @default.
- W2920293892 cites W2768267830 @default.
- W2920293892 cites W2890599985 @default.
- W2920293892 cites W2904142648 @default.
- W2920293892 cites W2950220847 @default.
- W2920293892 cites W2950928354 @default.
- W2920293892 cites W2950943852 @default.
- W2920293892 cites W2952062734 @default.
- W2920293892 cites W2963069632 @default.
- W2920293892 cites W2963446712 @default.
- W2920293892 cites W2963739978 @default.
- W2920293892 cites W2964160102 @default.
- W2920293892 cites W2964308564 @default.
- W2920293892 hasPublicationYear "2019" @default.
- W2920293892 type Work @default.
- W2920293892 sameAs 2920293892 @default.
- W2920293892 citedByCount "7" @default.
- W2920293892 countsByYear W29202938922019 @default.
- W2920293892 countsByYear W29202938922020 @default.
- W2920293892 countsByYear W29202938922021 @default.
- W2920293892 crossrefType "posted-content" @default.
- W2920293892 hasAuthorship W2920293892A5044802273 @default.
- W2920293892 hasAuthorship W2920293892A5070990160 @default.
- W2920293892 hasAuthorship W2920293892A5078600050 @default.
- W2920293892 hasAuthorship W2920293892A5080304958 @default.
- W2920293892 hasAuthorship W2920293892A5087695030 @default.
- W2920293892 hasConcept C11413529 @default.
- W2920293892 hasConcept C118615104 @default.
- W2920293892 hasConcept C121332964 @default.
- W2920293892 hasConcept C12426560 @default.
- W2920293892 hasConcept C134306372 @default.
- W2920293892 hasConcept C154945302 @default.
- W2920293892 hasConcept C177148314 @default.
- W2920293892 hasConcept C186633575 @default.
- W2920293892 hasConcept C190470478 @default.
- W2920293892 hasConcept C202444582 @default.
- W2920293892 hasConcept C2524010 @default.
- W2920293892 hasConcept C26405456 @default.
- W2920293892 hasConcept C2778530986 @default.
- W2920293892 hasConcept C33923547 @default.
- W2920293892 hasConcept C37914503 @default.
- W2920293892 hasConcept C41008148 @default.
- W2920293892 hasConcept C50644808 @default.
- W2920293892 hasConcept C62520636 @default.
- W2920293892 hasConceptScore W2920293892C11413529 @default.
- W2920293892 hasConceptScore W2920293892C118615104 @default.
- W2920293892 hasConceptScore W2920293892C121332964 @default.
- W2920293892 hasConceptScore W2920293892C12426560 @default.
- W2920293892 hasConceptScore W2920293892C134306372 @default.
- W2920293892 hasConceptScore W2920293892C154945302 @default.
- W2920293892 hasConceptScore W2920293892C177148314 @default.
- W2920293892 hasConceptScore W2920293892C186633575 @default.
- W2920293892 hasConceptScore W2920293892C190470478 @default.
- W2920293892 hasConceptScore W2920293892C202444582 @default.
- W2920293892 hasConceptScore W2920293892C2524010 @default.
- W2920293892 hasConceptScore W2920293892C26405456 @default.
- W2920293892 hasConceptScore W2920293892C2778530986 @default.
- W2920293892 hasConceptScore W2920293892C33923547 @default.
- W2920293892 hasConceptScore W2920293892C37914503 @default.
- W2920293892 hasConceptScore W2920293892C41008148 @default.
- W2920293892 hasConceptScore W2920293892C50644808 @default.
- W2920293892 hasConceptScore W2920293892C62520636 @default.
- W2920293892 hasLocation W29202938921 @default.
- W2920293892 hasOpenAccess W2920293892 @default.
- W2920293892 hasPrimaryLocation W29202938921 @default.
- W2920293892 hasRelatedWork W1133641091 @default.
- W2920293892 hasRelatedWork W1572736436 @default.
- W2920293892 hasRelatedWork W183895133 @default.
- W2920293892 hasRelatedWork W2144513243 @default.
- W2920293892 hasRelatedWork W2279639021 @default.
- W2920293892 hasRelatedWork W2344685289 @default.
- W2920293892 hasRelatedWork W2605372163 @default.
- W2920293892 hasRelatedWork W2777256551 @default.
- W2920293892 hasRelatedWork W2883570905 @default.
- W2920293892 hasRelatedWork W2904281284 @default.
- W2920293892 hasRelatedWork W2914978357 @default.
- W2920293892 hasRelatedWork W2937006657 @default.
- W2920293892 hasRelatedWork W2949965765 @default.
- W2920293892 hasRelatedWork W2962933129 @default.
- W2920293892 hasRelatedWork W2963000077 @default.
- W2920293892 hasRelatedWork W2990086510 @default.
- W2920293892 hasRelatedWork W3096695297 @default.
- W2920293892 hasRelatedWork W3098754584 @default.