New method predicts compression errors in neural representations without ground truth
Non-asymptotic bounds for INRs, hash encodings, and tensor factorizations — no original signal needed.
A new paper from Quang Luong Nhat Nguyen and Sara Fridovich-Keil tackles a fundamental problem in computational imaging: how to evaluate reconstruction fidelity when the original (ground truth) signal is unavailable. Their framework, published on arXiv (2606.00126), works for differentiable signal parameterizations such as implicit neural representations (INRs) and hybrid models that are increasingly used in neural rendering, medical imaging, and inverse problems.
The key insight is that when compression satisfies certain natural properties, the error at any compression level is bounded by a simple scaled difference between predictions at two different compression levels. These non-asymptotic, signal-specific bounds are both theoretically sound and efficiently computable. The authors verify the required properties for interpolated grids, Fourier feature networks, multi-resolution hash encodings, and tensor factorizations. Empirically, the method tracks global error curves and produces informative local error heatmaps for synthetic signals, natural images, radiance fields (NeRF-style), and MRI reconstruction — all without needing the ground truth. Code is available on GitHub.
- Proves non-asymptotic bounds on compression error for INRs, Fourier feature networks, hash encodings, and tensor factorizations
- Bounds are computed using only model outputs at different compression levels — no ground truth required
- Validated on radiance field reconstruction and MRI, producing tight error curves and local heatmaps
Why It Matters
Enables trustworthy quality assessment in neural rendering and medical imaging when original data is inaccessible.