Extension

rmath.nn

Neural network building blocks: activation functions, loss functions, normalization layers, and pooling. All operate on rmath.Array.

Activation Functions

Function	Description
`relu(a)`	Rectified linear unit: max(0, x)
`leaky_relu(a, alpha)`	Leaky ReLU with slope alpha for negative inputs
`sigmoid(a)`	Logistic sigmoid: 1 / (1 + e⁻ˣ)
`gelu(a)`	Gaussian error linear unit
`softmax(a)`	Row-wise softmax (numerically stable)
`hardswish(a)`	Hard-swish activation

Loss Functions

Function	Returns	Description
`mse_loss(a, target)`	float	Mean squared error
`cross_entropy_loss(a, labels)`	float	Cross-entropy loss. Labels are integer class indices.

Normalization & Regularization

Function	Description
`batch_norm(a, mu, sigma, gamma, beta)`	Batch normalization. mu, sigma, gamma, beta are Vectors.
`layer_norm(a, eps)`	Layer normalization with epsilon for numerical stability.
`dropout(a, p)`	Random dropout with probability p.

Pooling

`max_pool2d(a, kernel_size)`	Array	2D max pooling
`avg_pool2d(a, kernel_size)`	Array	2D average pooling