Struct NaturalGradientParams 

pub struct NaturalGradientParams {
    pub eta: Array1<f64>,
    pub fisher_diag: Array1<f64>,
}

Natural gradient parameters for exponential family distributions

Fields

eta: Array1<f64>

Natural parameters (eta) for the variational distribution

fisher_diag: Array1<f64>

Fisher information matrix (or its approximation) Stored as diagonal for mean-field case

Implementations

impl NaturalGradientParams

pub fn from_mean_field(mf: &MeanFieldGaussian) -> Self

Create natural gradient parameters for a mean-field Gaussian

For a Gaussian q(z; mu, sigma^2): eta_1 = mu / sigma^2 (natural parameter 1) eta_2 = -1 / (2*sigma^2) (natural parameter 2)

Fisher information is 1/sigma^2 for the mean parameter and 2/sigma^4 for the variance parameter

pub fn to_mean_field(&self) -> Result<MeanFieldGaussian>

Convert natural parameters back to mean/std parameterization

pub fn natural_gradient_update( &self, euclidean_grad: &Array1<f64>, ) -> Result<Array1<f64>>

Compute natural gradient update: update = Fisher^{-1} * euclidean_grad For diagonal Fisher, this is element-wise division

Trait Implementations

impl Clone for NaturalGradientParams

fn clone(&self) -> NaturalGradientParams

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for NaturalGradientParams

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.