Matches in SemOpenAlex for { <https://semopenalex.org/work/W3206270569> ?p ?o ?g. }
- W3206270569 abstract "We study the optimization problem associated with fitting two-layer ReLU neural networks with respect to the squared loss, where labels are generated by a target network. We make use of the rich symmetry structure to develop a novel set of tools for studying families of spurious minima. In contrast to existing approaches which operate in limiting regimes, our technique directly addresses the nonconvex loss landscape for a finite number of inputs $d$ and neurons $k$, and provides analytic, rather than heuristic, information. In particular, we derive analytic estimates for the loss at different minima, and prove that modulo $O(d^{-1/2})$-terms the Hessian spectrum concentrates near small positive constants, with the exception of $Theta(d)$ eigenvalues which grow linearly with~$d$. We further show that the Hessian spectrum at global and spurious minima coincide to $O(d^{-1/2})$-order, thus challenging our ability to argue about statistical generalization through local curvature. Lastly, our technique provides the exact emph{fractional} dimensionality at which families of critical points turn from saddles into spurious minima. This makes possible the study of the creation and the annihilation of spurious minima using powerful tools from equivariant bifurcation theory." @default.
- W3206270569 created "2021-10-25" @default.
- W3206270569 creator A5039879965 @default.
- W3206270569 creator A5044800354 @default.
- W3206270569 date "2021-07-21" @default.
- W3206270569 modified "2023-09-27" @default.
- W3206270569 title "Analytic Study of Families of Spurious Minima in Two-Layer ReLU Neural Networks: A Tale of Symmetry II" @default.
- W3206270569 cites W121410702 @default.
- W3206270569 cites W1533861849 @default.
- W3206270569 cites W196871588 @default.
- W3206270569 cites W1995842804 @default.
- W3206270569 cites W2006997461 @default.
- W3206270569 cites W2037985840 @default.
- W3206270569 cites W2042318263 @default.
- W3206270569 cites W2066424095 @default.
- W3206270569 cites W2070702809 @default.
- W3206270569 cites W2090614046 @default.
- W3206270569 cites W2093584183 @default.
- W3206270569 cites W2110166586 @default.
- W3206270569 cites W2150872430 @default.
- W3206270569 cites W2552194003 @default.
- W3206270569 cites W2566505556 @default.
- W3206270569 cites W2591714514 @default.
- W3206270569 cites W2605372163 @default.
- W3206270569 cites W2626325961 @default.
- W3206270569 cites W2731468224 @default.
- W3206270569 cites W27434444 @default.
- W3206270569 cites W2752366553 @default.
- W3206270569 cites W2752851182 @default.
- W3206270569 cites W2763374915 @default.
- W3206270569 cites W2768267830 @default.
- W3206270569 cites W2785533664 @default.
- W3206270569 cites W2809090039 @default.
- W3206270569 cites W2889737445 @default.
- W3206270569 cites W2904243021 @default.
- W3206270569 cites W2912173254 @default.
- W3206270569 cites W2912713668 @default.
- W3206270569 cites W2914484425 @default.
- W3206270569 cites W2947654433 @default.
- W3206270569 cites W2962939986 @default.
- W3206270569 cites W2963095610 @default.
- W3206270569 cites W2963211922 @default.
- W3206270569 cites W2963325933 @default.
- W3206270569 cites W2963383839 @default.
- W3206270569 cites W2963470399 @default.
- W3206270569 cites W2963519230 @default.
- W3206270569 cites W2963623651 @default.
- W3206270569 cites W2963650649 @default.
- W3206270569 cites W2963744427 @default.
- W3206270569 cites W2963959597 @default.
- W3206270569 cites W2964346549 @default.
- W3206270569 cites W2970618525 @default.
- W3206270569 cites W2981298641 @default.
- W3206270569 cites W3028642772 @default.
- W3206270569 cites W3037180580 @default.
- W3206270569 cites W3037508544 @default.
- W3206270569 cites W3089908128 @default.
- W3206270569 cites W3099497569 @default.
- W3206270569 cites W3100762814 @default.
- W3206270569 cites W3113714439 @default.
- W3206270569 cites W3135610712 @default.
- W3206270569 cites W3172370861 @default.
- W3206270569 cites W3180642516 @default.
- W3206270569 cites W3197486621 @default.
- W3206270569 cites W611802393 @default.
- W3206270569 cites W3141350557 @default.
- W3206270569 hasPublicationYear "2021" @default.
- W3206270569 type Work @default.
- W3206270569 sameAs 3206270569 @default.
- W3206270569 citedByCount "0" @default.
- W3206270569 crossrefType "posted-content" @default.
- W3206270569 hasAuthorship W3206270569A5039879965 @default.
- W3206270569 hasAuthorship W3206270569A5044800354 @default.
- W3206270569 hasConcept C105795698 @default.
- W3206270569 hasConcept C111030470 @default.
- W3206270569 hasConcept C121332964 @default.
- W3206270569 hasConcept C121864883 @default.
- W3206270569 hasConcept C134306372 @default.
- W3206270569 hasConcept C154945302 @default.
- W3206270569 hasConcept C158693339 @default.
- W3206270569 hasConcept C171036898 @default.
- W3206270569 hasConcept C186633575 @default.
- W3206270569 hasConcept C195065555 @default.
- W3206270569 hasConcept C202444582 @default.
- W3206270569 hasConcept C203616005 @default.
- W3206270569 hasConcept C2524010 @default.
- W3206270569 hasConcept C2779886137 @default.
- W3206270569 hasConcept C28826006 @default.
- W3206270569 hasConcept C33923547 @default.
- W3206270569 hasConcept C41008148 @default.
- W3206270569 hasConcept C50644808 @default.
- W3206270569 hasConcept C62520636 @default.
- W3206270569 hasConcept C97256817 @default.
- W3206270569 hasConceptScore W3206270569C105795698 @default.
- W3206270569 hasConceptScore W3206270569C111030470 @default.
- W3206270569 hasConceptScore W3206270569C121332964 @default.
- W3206270569 hasConceptScore W3206270569C121864883 @default.
- W3206270569 hasConceptScore W3206270569C134306372 @default.
- W3206270569 hasConceptScore W3206270569C154945302 @default.
- W3206270569 hasConceptScore W3206270569C158693339 @default.