Sketch-Based Regularization Enhances In Situ Training of Implicit Neural Compressors
A recent paper on arXiv (ID: 2511.02659) presents a novel protocol for in situ training of implicit neural representations (INRs). This approach utilizes limited memory buffers containing both full and sketched data samples to mitigate catastrophic forgetting. The theoretical foundation for employing sketching as a regularizer is derived from a Johnson-Lindenstrauss-informed result. The method emphasizes in situ neural compression through INR-based hypernetworks, tested on intricate simulation data in two and three dimensions, over extended time periods, and across unstructured grids and non-Cartesian geometries. Findings reveal impressive reconstruction capabilities at high compression rates, with sketching closely approximating full-data training performance. Authored by unnamed researchers, this work may influence continual learning, though it primarily addresses scientific simulation data compression.
Key facts
- Paper ID: arXiv:2511.02659
- Announcement type: replace-cross
- Method: in situ training protocol for implicit neural representations
- Uses limited memory buffers of full and sketched data samples
- Sketching prevents catastrophic forgetting
- Theoretical motivation: Johnson-Lindenstrauss-informed result
- Target: in situ neural compression using INR-based hypernetworks
- Evaluated on 2D and 3D simulation data, long time horizons, unstructured grids, non-Cartesian geometries
- Results: strong reconstruction at high compression rates
- Sketching allows approximate match to full-data performance
Entities
Institutions
- arXiv