Skip to content

Structured reservoir computing #285

@MartinuzziFrancesco

Description

@MartinuzziFrancesco

Description

Introduced in Dong2020 this initialization uses structured transforms to speed-up computations of random matrix multiplications. Also based on Yu2016

Following the description in Dong2020: The general idea is to replace the dense matrices $W = [W, W_{\text{in}}]$ by a succession of Hadamard matrices $H$ (structured orthonormal matrices composed of $\pm1/\sqrt{p}$ components) and diagonal random matrices $D_i$ for $i = 1, 2, 3$ sampled from an i.i.d. Rademacher distribution:

$$ W = \frac{\sqrt{p}}{\sigma} H D_1 H D_2 H D_3 \tag{10} $$

Original implementation

Structured init in the test code provided with the paper

Metadata

Metadata

Assignees

No one assigned

    Labels

    initsinitializers for internal matrices

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions