Fix TSNE gradient computation

[andrewdalpino]

t-SNE gradient is mathematically wrong — src/Transformers/TSNE.php:523-543 The q-kernel uses raw (non-squared) Euclidean distances (Euclidean::compute() returns sqrt(Σ(a−b)²)), the gradient weight is raw d instead of (1 + d²/ν)^⁻¹, and Q is normalized by sum()×2 including the diagonal (all 1.0). The embedding does not minimize the documented KL divergence; wrong output is produced silently.

[Copilot]

Pull request overview

This PR aims to correct the mathematical implementation of the t-SNE gradient computation so that optimization steps better align with the intended KL divergence objective.

Changes:

• Update the low-dimensional Student-t kernel to use squared distances and adjust Q normalization to exclude the diagonal.
• Add a targeted unit test for the gradient() calculation.
• Document the fix in the changelog.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
src/Transformers/TSNE.php	Fixes kernel/normalization pieces used in gradient computation.
tests/Transformers/TSNETest.php	Adds a regression test for gradient behavior (and helper to invoke it).
CHANGELOG.md	Records the fix under a new 2.5.3 entry.

---

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

[apphp]

Very good fix!