Matches in SemOpenAlex for { <https://semopenalex.org/work/W4386076402> ?p ?o ?g. }
- W4386076402 abstract "Masked autoencoder (MAE), a simple and effective self-supervised learning framework based on the reconstruction of masked image regions, has recently achieved prominent success in a variety of vision tasks. Despite the emergence of intriguing empirical observations on MAE, a theoretically principled understanding is still lacking. In this work, we formally characterize and justify existing empirical in-sights and provide theoretical guarantees of MAE. We formulate the underlying data-generating process as a hierarchical latent variable model, and show that under reasonable assumptions, MAE provably identifies a set of latent variables in the hierarchical model, explaining why MAE can extract high-level information from pixels. Further, we show how key hyperparameters in MAE (the masking ratio and the patch size) determine which true latent variables to be recovered, therefore influencing the level of semantic information in the representation. Specifically, extremely large or small masking ratios inevitably lead to low-level representations. Our theory offers coherent explanations of existing empirical observations and provides insights for potential empirical improvements and fundamental limitations of the masked-reconstruction paradigm. We conduct extensive experiments to validate our theoretical insights." @default.
- W4386076402 created "2023-08-23" @default.
- W4386076402 creator A5009547049 @default.
- W4386076402 creator A5016038041 @default.
- W4386076402 creator A5053809095 @default.
- W4386076402 creator A5060335438 @default.
- W4386076402 creator A5078074297 @default.
- W4386076402 creator A5081398601 @default.
- W4386076402 creator A5088257057 @default.
- W4386076402 date "2023-06-01" @default.
- W4386076402 modified "2023-10-18" @default.
- W4386076402 title "Understanding Masked Autoencoders via Hierarchical Latent Variable Models" @default.
- W4386076402 cites W2099741732 @default.
- W4386076402 cites W2108384452 @default.
- W4386076402 cites W2133665775 @default.
- W4386076402 cites W2141983208 @default.
- W4386076402 cites W2565639579 @default.
- W4386076402 cites W2919234133 @default.
- W4386076402 cites W2962770929 @default.
- W4386076402 cites W2963150697 @default.
- W4386076402 cites W3035524453 @default.
- W4386076402 cites W3145450063 @default.
- W4386076402 cites W3159481202 @default.
- W4386076402 cites W3177096435 @default.
- W4386076402 cites W4205778870 @default.
- W4386076402 cites W4312804044 @default.
- W4386076402 doi "https://doi.org/10.1109/cvpr52729.2023.00765" @default.
- W4386076402 hasPublicationYear "2023" @default.
- W4386076402 type Work @default.
- W4386076402 citedByCount "0" @default.
- W4386076402 crossrefType "proceedings-article" @default.
- W4386076402 hasAuthorship W4386076402A5009547049 @default.
- W4386076402 hasAuthorship W4386076402A5016038041 @default.
- W4386076402 hasAuthorship W4386076402A5053809095 @default.
- W4386076402 hasAuthorship W4386076402A5060335438 @default.
- W4386076402 hasAuthorship W4386076402A5078074297 @default.
- W4386076402 hasAuthorship W4386076402A5081398601 @default.
- W4386076402 hasAuthorship W4386076402A5088257057 @default.
- W4386076402 hasConcept C101738243 @default.
- W4386076402 hasConcept C108583219 @default.
- W4386076402 hasConcept C112933361 @default.
- W4386076402 hasConcept C119857082 @default.
- W4386076402 hasConcept C134306372 @default.
- W4386076402 hasConcept C136197465 @default.
- W4386076402 hasConcept C142362112 @default.
- W4386076402 hasConcept C153349607 @default.
- W4386076402 hasConcept C154945302 @default.
- W4386076402 hasConcept C177264268 @default.
- W4386076402 hasConcept C17744445 @default.
- W4386076402 hasConcept C182365436 @default.
- W4386076402 hasConcept C199360897 @default.
- W4386076402 hasConcept C199539241 @default.
- W4386076402 hasConcept C26517878 @default.
- W4386076402 hasConcept C2776359362 @default.
- W4386076402 hasConcept C2777402240 @default.
- W4386076402 hasConcept C33923547 @default.
- W4386076402 hasConcept C38652104 @default.
- W4386076402 hasConcept C41008148 @default.
- W4386076402 hasConcept C51167844 @default.
- W4386076402 hasConcept C65965080 @default.
- W4386076402 hasConcept C8642999 @default.
- W4386076402 hasConcept C94625758 @default.
- W4386076402 hasConceptScore W4386076402C101738243 @default.
- W4386076402 hasConceptScore W4386076402C108583219 @default.
- W4386076402 hasConceptScore W4386076402C112933361 @default.
- W4386076402 hasConceptScore W4386076402C119857082 @default.
- W4386076402 hasConceptScore W4386076402C134306372 @default.
- W4386076402 hasConceptScore W4386076402C136197465 @default.
- W4386076402 hasConceptScore W4386076402C142362112 @default.
- W4386076402 hasConceptScore W4386076402C153349607 @default.
- W4386076402 hasConceptScore W4386076402C154945302 @default.
- W4386076402 hasConceptScore W4386076402C177264268 @default.
- W4386076402 hasConceptScore W4386076402C17744445 @default.
- W4386076402 hasConceptScore W4386076402C182365436 @default.
- W4386076402 hasConceptScore W4386076402C199360897 @default.
- W4386076402 hasConceptScore W4386076402C199539241 @default.
- W4386076402 hasConceptScore W4386076402C26517878 @default.
- W4386076402 hasConceptScore W4386076402C2776359362 @default.
- W4386076402 hasConceptScore W4386076402C2777402240 @default.
- W4386076402 hasConceptScore W4386076402C33923547 @default.
- W4386076402 hasConceptScore W4386076402C38652104 @default.
- W4386076402 hasConceptScore W4386076402C41008148 @default.
- W4386076402 hasConceptScore W4386076402C51167844 @default.
- W4386076402 hasConceptScore W4386076402C65965080 @default.
- W4386076402 hasConceptScore W4386076402C8642999 @default.
- W4386076402 hasConceptScore W4386076402C94625758 @default.
- W4386076402 hasFunder F4320306076 @default.
- W4386076402 hasFunder F4320337345 @default.
- W4386076402 hasLocation W43860764021 @default.
- W4386076402 hasOpenAccess W4386076402 @default.
- W4386076402 hasPrimaryLocation W43860764021 @default.
- W4386076402 hasRelatedWork W1525474987 @default.
- W4386076402 hasRelatedWork W2770818364 @default.
- W4386076402 hasRelatedWork W2786772298 @default.
- W4386076402 hasRelatedWork W2905986396 @default.
- W4386076402 hasRelatedWork W2933374552 @default.
- W4386076402 hasRelatedWork W2952508194 @default.
- W4386076402 hasRelatedWork W3123307809 @default.
- W4386076402 hasRelatedWork W4285137313 @default.
- W4386076402 hasRelatedWork W4380136218 @default.