Reproduce “RECOMBINER: Robust and Enhanced Compression with Bayesian Implicit Neural Representations” 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
Jiajun He, Gergely Flamich, Zongyu Guo, José Miguel Hernández-Lobato
RECOMBINER, developed by researchers from the University of Cambridge and USTC, enhances implicit neural representation (INR) based data compression by improving variational approximations, enabling local feature adaptation, and scaling to high-resolution data. It achieves competitive performance across images, audio, video, and 3D protein structures, outperforming its predecessor COMBINER and even VAE-based codecs on low-resolution images at low bitrates.
Key insights
- Linear weight reparameterization allows for an expressive, implicit full-covariance posterior over INR weights, enabling better data fitting and efficient pruning.
- Learnable positional encodings provide a mechanism for INRs to adapt to local data details, improving reconstruction quality and stabilizing optimization, especially for fine-grained information.
- A hierarchical Bayesian model combined with patching and adaptive bit allocation via permutation allows RECOMBINER to scale effectively to high-resolution data while managing inter-patch dependencies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.