`cuML` Barnes-Hut t-SNE collapses to a straight line on homogeneous embeddings

[NetZissou]

cuML’s t-SNE supports three algorithms: the original exact algorithm (O(N^2)), the Barnes-Hut approximation and the fast Fourier transform interpolation approximation (O(N log N)). The latter two are derived from CannyLabs’ open-source CUDA code and produce extremely fast embeddings when n_components = 2. The exact algorithm is more accurate, but too slow to use on large datasets.

In the embed_explore / precalculated apps, t-SNE projection with the cuML (GPU) backend renders as a straight 45° line for the Darwin's-finches BioCLIP 2 embeddings. PCA & UMAP are fine, and switching the backend to sklearn produces a correct t-SNE, therefore this is specific to cuML's Barnes-Hut t-SNE, not the data or our pipeline.

The app uses cuML t-SNE's default method='barnes_hut', which collapses both output dimensions onto one axis on this data. method='exact' fixes it. Barnes-Hut's degeneracy is data-dependent. The finch embeddings are extremely homogeneous, near-uniform pairwise structure after L2-normalization.

Degenerate / collapsed embedding: 2D output isn't a real spread but lies on a single line (one output axis carries ~all the variance; the two coordinates become perfectly correlated)

cuML t-SNE result

cuML PCA result

cuML UMAP result

sklearn t-SNE result