A new arXiv preprint examines how Gaussian kernel similarity measures can be computed more efficiently for large sets of points. The authors build on Random Fourier Features to construct low-dimensional embeddings tailored to data lying on manifolds. The work targets applications such as kernel PCA and spectral clustering, where pairwise kernel evaluations are typically costly.
Researchers introduce MAPLE, a nonlinear dimensionality reduction technique that builds on UMAP by adding a self-supervised learning component to better capture manifold structure. The method aims to encode low-dimensional manifold geometry more efficiently, supporting visual analysis tasks. The work is described in an arXiv preprint in machine learning.
A new arXiv paper proposes pairing an energy-based model with an autoencoder in a cooperative training setup. The method alternates between refining the learned energy landscape and projecting samples back onto a data manifold, aiming to overcome common training difficulties in energy-based generative models. The authors report the framework assigns low energy to realistic samples and higher energy to unlikely ones.
A machine learning preprint on arXiv introduces a generative transport method aimed at datasets that concentrate near a low-dimensional structure embedded in a high-dimensional space. The approach seeks to limit probability mass leaking away from the data-supporting manifold while staying computationally practical, in contrast to the iterative sampling used by diffusion models. The paper was updated as a v2 cross-list replacement.