This paper introduces Error-based Context Pruning (ECoP), an efficient meta-learning technique for optimizing large-scale implicit neural representations (INRs). ECoP addresses the memory limitations of existing optimization-based meta-learning methods by selectively choosing a subset of context points at each adaptation step based on predictive error. This adaptive selection prioritizes global structure initially and then high-frequency details, leading to improved reconstruction and enabling learning on high-dimensional signals. The method incorporates bootstrapped correction to mitigate information loss from pruning and uses gradient scaling at test time to improve performance. ECoP is model-agnostic and demonstrates significant improvements across various signal modalities.

Jihoon Tack, Subin Kim, Sihyun Yu, Jaeho Lee, Jinwoo Shin, Jonathan Richard Schwarz