Fast determinantal sampling on general spaces and diffusion geometry
arXiv:2607.06644
2026
Training
2 ideas extracted · analyzed Aug 30, 2026
What the math gives to ML
The paper provides a mathematically principled way to construct minibatches using determinantal point processes whose geometry is defined by a graph or manifold diffusion operator rather than ambient Euclidean distance. The transferable asset is the combination of repulsive sampling, diffusion-Laplacian eigenfunctions, and uniform Dirichlet-energy and norm-approximation guarantees that depend on intrinsic dimension. The most direct neural-network use is a fixed-size DPP sampler over training examples, with a diffusion kernel built from an approximate k-nearest-neighbor graph and inclusion-probability weighting for unbiased gradient estimates. A second direction is to use the paper's energy and almost-isometry bounds as an adaptive diagnostic for whether a minibatch preserves low-frequency loss or gradient structure.
Ideas from this paper
✗ Failed on benchmark
2026
Replace uniform minibatch sampling by a fixed-size determinantal point process whose similarity matrix is a diffusion kernel on the training-data k-NN graph. The sampler repels nearby or redundant examples while preserving multiple diffusion modes, so a small batch should cover intrinsic data geometry and provide lower-variance estimates of losses and gradients.
Useful8/10
Difficulty6/10
Novelty5/10
Unverified
2026
Use the diffusion graph's Dirichlet energy and almost-isometry inequalities to score whether a candidate minibatch preserves the low-frequency structure of losses, logits, or gradients over the dataset. Reject or augment batches that distort these quantities, producing a geometry-aware batch acceptance rule rather than relying only on random or loss-based sampling.
Useful6/10
Difficulty7/10
Novelty7/10