Robust Differentiable SVD

Eigendecomposition of symmetric matrices is at the heart of many computer vision algorithms. However, the derivatives of the eigenvectors tend to be numerically unstable, whether using the SVD to compute them analytically or using the Power Iteration (PI) method to approximate them. This instability...

Ausführliche Beschreibung

Bibliographische Detailangaben
Veröffentlicht in:IEEE transactions on pattern analysis and machine intelligence. - 1979. - 44(2022), 9 vom: 03. Sept., Seite 5472-5487
1. Verfasser: Wang, Wei (VerfasserIn)
Weitere Verfasser: Dang, Zheng, Hu, Yinlin, Fua, Pascal, Salzmann, Mathieu
Format: Online-Aufsatz
Sprache:English
Veröffentlicht: 2022
Zugriff auf das übergeordnete Werk:IEEE transactions on pattern analysis and machine intelligence
Schlagworte:Journal Article
LEADER 01000naa a22002652 4500
001 NLM323996620
003 DE-627
005 20231225185354.0
007 cr uuu---uuuuu
008 231225s2022 xx |||||o 00| ||eng c
024 7 |a 10.1109/TPAMI.2021.3072422  |2 doi 
028 5 2 |a pubmed24n1079.xml 
035 |a (DE-627)NLM323996620 
035 |a (NLM)33844626 
040 |a DE-627  |b ger  |c DE-627  |e rakwb 
041 |a eng 
100 1 |a Wang, Wei  |e verfasserin  |4 aut 
245 1 0 |a Robust Differentiable SVD 
264 1 |c 2022 
336 |a Text  |b txt  |2 rdacontent 
337 |a ƒaComputermedien  |b c  |2 rdamedia 
338 |a ƒa Online-Ressource  |b cr  |2 rdacarrier 
500 |a Date Revised 05.08.2022 
500 |a published: Print-Electronic 
500 |a Citation Status PubMed-not-MEDLINE 
520 |a Eigendecomposition of symmetric matrices is at the heart of many computer vision algorithms. However, the derivatives of the eigenvectors tend to be numerically unstable, whether using the SVD to compute them analytically or using the Power Iteration (PI) method to approximate them. This instability arises in the presence of eigenvalues that are close to each other. This makes integrating eigendecomposition into deep networks difficult and often results in poor convergence, particularly when dealing with large matrices. While this can be mitigated by partitioning the data into small arbitrary groups, doing so has no theoretical basis and makes it impossible to exploit the full power of eigendecomposition. In previous work, we mitigated this using SVD during the forward pass and PI to compute the gradients during the backward pass. However, the iterative deflation procedure required to compute multiple eigenvectors using PI tends to accumulate errors and yield inaccurate gradients. Here, we show that the Taylor expansion of the SVD gradient is theoretically equivalent to the gradient obtained using PI without relying in practice on an iterative process and thus yields more accurate gradients. We demonstrate the benefits of this increased accuracy for image classification and style transfer 
650 4 |a Journal Article 
700 1 |a Dang, Zheng  |e verfasserin  |4 aut 
700 1 |a Hu, Yinlin  |e verfasserin  |4 aut 
700 1 |a Fua, Pascal  |e verfasserin  |4 aut 
700 1 |a Salzmann, Mathieu  |e verfasserin  |4 aut 
773 0 8 |i Enthalten in  |t IEEE transactions on pattern analysis and machine intelligence  |d 1979  |g 44(2022), 9 vom: 03. Sept., Seite 5472-5487  |w (DE-627)NLM098212257  |x 1939-3539  |7 nnns 
773 1 8 |g volume:44  |g year:2022  |g number:9  |g day:03  |g month:09  |g pages:5472-5487 
856 4 0 |u http://dx.doi.org/10.1109/TPAMI.2021.3072422  |3 Volltext 
912 |a GBV_USEFLAG_A 
912 |a SYSFLAG_A 
912 |a GBV_NLM 
912 |a GBV_ILN_350 
951 |a AR 
952 |d 44  |j 2022  |e 9  |b 03  |c 09  |h 5472-5487