Struct BayesianNeuralNetwork 

pub struct BayesianNeuralNetwork<F> {
    pub architecture: Vec<usize>,
    pub activations: Vec<ActivationType>,
    pub weight_priors: Vec<DistributionType<F>>,
    pub bias_priors: Vec<DistributionType<F>>,
    pub weight_samples: Option<Vec<Vec<Array2<F>>>>,
    pub bias_samples: Option<Vec<Vec<Array1<F>>>>,
}

Bayesian neural network implementation

Fields

architecture: Vec<usize>

Network architecture

activations: Vec<ActivationType>

Activation functions per layer

weight_priors: Vec<DistributionType<F>>

Weight priors

bias_priors: Vec<DistributionType<F>>

Bias priors

weight_samples: Option<Vec<Vec<Array2<F>>>>

Trained posterior ensemble of weights: weight_samples[m][l] is the weight matrix of layer l for ensemble member m. Populated by BayesianNeuralNetwork::fit; None until then.

bias_samples: Option<Vec<Vec<Array1<F>>>>

Trained posterior ensemble of biases: bias_samples[m][l] is the bias vector of layer l for ensemble member m. Populated by BayesianNeuralNetwork::fit; None until then.

Implementations

impl<F: AdvancedBayesianFloat> BayesianNeuralNetwork<F>

pub fn fit( &mut self, x: &ArrayView2<'_, F>, y: &ArrayView2<'_, F>, config: &BnnTrainingConfig, ) -> StatsResult<()>

Train a deep ensemble of posterior weight/bias samples: each member is randomly initialized from the network’s priors, optionally bootstrap-resampled, and trained to a local optimum of the mean squared error via exact backpropagation + gradient descent. Populates self.weight_samples / self.bias_samples, which Self::predict_with_uncertainty then draws from.

pub fn predict_with_uncertainty( &self, x: &ArrayView2<'_, F>, n_samples_: usize, ) -> StatsResult<(Array2<F>, Array2<F>)>

Make predictions with uncertainty quantification.

If Self::fit has already trained a posterior ensemble, the ensemble is a small finite population (typically a handful to a few dozen members) whose exact predictive mean/variance is directly computable, so this forward-propagates through every trained member exactly once (n_samples_ is ignored in this case: randomly resampling a finite, fully-known population with replacement would only add spurious Monte Carlo noise around a quantity that has no randomness left to estimate). Otherwise – no ensemble has been trained yet – this falls back to real prior-predictive Monte Carlo: n_samples_ independent draws of the whole network from its priors, each forward-propagated through the actual input x, where genuine random sampling is unavoidable since the prior is a continuous distribution. Either way, the reported mean/variance are genuine empirical statistics of real forward passes – never a constant placeholder.

impl<F: AdvancedBayesianFloat> BayesianNeuralNetwork<F>

pub fn new( architecture: Vec<usize>, activations: Vec<ActivationType>, ) -> StatsResult<Self>

Create new Bayesian neural network

pub fn forward( &self, x: &ArrayView2<'_, F>, weights: &[Array2<F>], biases: &[Array1<F>], ) -> StatsResult<Array2<F>>

Forward pass through the network

Trait Implementations

impl<F: Clone> Clone for BayesianNeuralNetwork<F>

fn clone(&self) -> BayesianNeuralNetwork<F>

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<F: Debug> Debug for BayesianNeuralNetwork<F>

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

Formats the value using the given formatter.