\(\mathrm{SU}(1,1)\) equivariant neural networks and application to robust Toeplitz Hermitian positive definite matrix classification (Q2117902)

From MaRDI portal
scientific article
Language Label Description Also known as
English
\(\mathrm{SU}(1,1)\) equivariant neural networks and application to robust Toeplitz Hermitian positive definite matrix classification
scientific article

    Statements

    \(\mathrm{SU}(1,1)\) equivariant neural networks and application to robust Toeplitz Hermitian positive definite matrix classification (English)
    0 references
    0 references
    0 references
    0 references
    22 March 2022
    0 references
    Deep learning architectures are very efficient in representing complex data. If the problem at hand has symmetries, special architectures are proposed. For instance, convolutional neural networks respect translational symmetry. Group-equivariant neural networks for other group symmetries is a hot topic, and recently they gained huge attention. Recently, for the success of AlphaFold, a system solving the 3D protein structure prediction, group-equivariant neural networks have been instrumental. Group-Convolutional Neural Networks (G-CNN) have been investigated by several authors. The robustness and data effiency of G-CNN are very important in real-world applications. The authors propose an alternative method for building \(\mathrm{SU}(1, 1)\)-equivariant neural networks on the Poincaré disk. In order to obtain good estimates of the integral, they define a Monte-Carlo sampling procedure on \(\mathrm{SU}(1,1)\) by using Helgason-Fourier (HF) transforms of functions defined on homogeneous spaces. As an application they consider classification of Toeplitz-Hermitian positive definite matrices, after a suitable identificiation. This kind of classification can be used in Doppler signal processing tasks. For the entire collection see [Zbl 1482.94007].
    0 references
    homogeneous spaces
    0 references
    Burg algorithm
    0 references
    radar clutter
    0 references
    equivariant neural networks
    0 references
    hyperbolic embedding
    0 references

    Identifiers

    0 references
    0 references
    0 references
    0 references
    0 references
    0 references