Fix Batch Norm

[andrewdalpino]

BatchNorm normalizes by width, not batch size — src/NeuralNet/Layers/BatchNorm.php:184, 290-293. Matrix::variance() divides by row count (features); backward uses dOut->m() where batch size n() belongs. Masked by square test matrices.

[Copilot]

Pull request overview

This PR fixes BatchNorm to normalize using the batch size (matrix n()) rather than the feature width (matrix m()), addressing incorrect statistics/gradients that were previously hidden by square (m == n) test matrices.

Changes:

• Update forward-pass variance calculation to use a mean-of-squared-deviations formulation aligned to the batch dimension.
• Fix backward gradient scaling to use dOut->n() (batch size) instead of dOut->m() (width/features).
• Add a regression test using a non-square input matrix to ensure the bug can’t be masked.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
tests/NeuralNet/Layers/BatchNormTest.php	Adds a non-square (3×4) regression test to validate batch-dimension normalization in forward/backward/infer.
src/NeuralNet/Layers/BatchNorm.php	Corrects variance computation and gradient scaling to normalize over batch size (n()).

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[apphp]

LGTM