Reproduce “RANDSMAPs: Random-Feature/multi-Scale Neural Decoders with Mass Preservation” 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
Dimitrios G. Patsatzis, Alessandro Della Pia, Lucia Russo, Constantinos Siettos
Researchers from Naples developed RANDSMAPs, a framework for neural decoders that reconstruct high-dimensional data from low-dimensional latent representations. These decoders explicitly guarantee the preservation of physical conservation laws, such as mass, by integrating numerical analysis principles and multi-scale random features, leading to accurate and computationally efficient results.
Key insights
- Traditional nonlinear decoders, such as Double Diffusion Maps (DDM), do not inherently preserve mass, which limits their applicability to data governed by conservation laws.
- By redesigning Random Feature Neural Networks to include an output bias column and training via constrained least-squares, exact mass preservation can be guaranteed for both training and unseen data.
- Multi-scale random features, derived from a hierarchical sampling of frequencies, effectively enhance the expressivity and approximation power of kernel-based decoders, improving their ability to model complex data geometries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.