Reproduce “Curriculum Learning for ab initio Deep Learned Refractive Optics” 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

Xinge Yang, Qiang Fu, Wolfgang Heidrich

This research introduces a curriculum learning framework for the ab initio design of compound refractive lenses, which allows for automated optimization starting from randomly initialized surfaces. The approach demonstrates enhanced stability and performance in creating complex optical systems, including a compact lens for extended depth-of-field imaging.

Key insights

  • A structured curriculum path, combined with regularization techniques, enables stable *ab initio* optimization of complex refractive lenses from random initial parameters, overcoming common convergence issues.
  • Dynamically focusing optimization on high-error image regions with a re-weighting mask is effective for navigating non-convex design spaces and avoiding suboptimal local minima.
  • The joint optimization of hybrid optical surfaces (incorporating odd polynomial terms) and a neural reconstruction network can produce compact, high-performance computational cameras for challenging specifications like extended depth-of-field.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.