Splitting the Difference Between Deep and Shallow Solutions of Inverse Problems

Explore a lecture on the intersection of compressed sensing and deep learning in computational imaging. Delve into Ben Recht's presentation at IPAM's Multi-Modal Imaging Workshop, where he discusses novel approaches to inverse problems. Learn about the Plenoxels system for photorealistic view synthesis, which offers a faster alternative to Neural Radiance Fields. Discover how sparse 3D grids with spherical harmonics can be optimized using gradient methods and convex regularization. Examine the trade-offs between theoretical guarantees and flexibility in imaging techniques, and understand the potential for combining the strengths of compressed sensing and deep learning paradigms. Gain insights into the challenges and advancements in multidimensional image processing, nonlinear measurements, and complex forward models in computational imaging.

Intro
Presentation
Sparse recovery
Deep nihilism
Problems with deep solutions
Novel view synthesis
Nerf papers
How is Nerf different
Rendering a scene
Volumetric formulation
Transmittance equation
Multiview synthesis
Neural Radiance
Problems with Neural Radiance
Planoxyls
Regularization
Resolution
Harmonic Basis
Results
TV regularization
Synthetic scenes
More scenes
The 360 view