Reproduce “Diffeomorphic Measure Matching with Kernels for Generative Modeling” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Biraj Pandey, Bamdad Hosseini, Pau Batlle, Houman Owhadi
Researchers developed KODE, a generative modeling framework that employs ordinary differential equations with Reproducing Kernel Hilbert Space (RKHS) parameterized velocity fields and Maximum Mean Discrepancy (MMD) to transport probability measures. The method offers theoretical error bounds and shows competitive performance against neural network-based flows on several benchmarks while requiring fewer parameters and faster training times in many cases.
Key insights
- RKHS-based velocity fields provide a principled way to construct diffeomorphic transport maps, inheriting smoothness and regularization properties from the field of diffeomorphic matching.
- The use of Maximum Mean Discrepancy (MMD) offers a robust, kernel-based metric for comparing probability distributions, suitable for generative modeling without explicit density knowledge.
- The theoretical framework provides quantitative error bounds, partitioning the MMD error into approximation error (related to inducing points), sample complexity, and model misspecification, which are typically scarce for generative models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.