Reproduce “Local Positional Encoding for Multi-Layer Perceptrons” 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
Shin Fujieda, Atsushi Yoshimura, Takahiro Harada
Researchers at AMD developed Local Positional Encoding (LPE), a novel input encoding method that enables small Multi-Layer Perceptrons to represent high-frequency signals efficiently. LPE achieves visually superior 2D image reconstruction and high-fidelity 3D Signed Distance Functions, outperforming standard positional and grid encodings while maintaining an equivalent memory footprint.
Key insights
- Localizing the effect of positional encoding to individual grid cells via learned amplitude modulation allows small MLPs to effectively capture high-frequency details.
- Replacing the lowest-frequency sinusoidal components with interpolated latent coefficients from grid vertices effectively mitigates boundary discontinuities inherent in localized encodings.
- The hybrid approach efficiently combines the high-frequency expressiveness of positional encoding with the localized feature representation and small MLP compatibility of grid encoding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.