NTK-Guided Implicit Neural Teaching

TL;DR

This paper introduces NINT, a method that accelerates training of implicit neural representations by selectively sampling coordinates based on NTK-guided loss gradients, reducing training time while maintaining quality.

Contribution

The paper proposes NTK-Guided Implicit Neural Teaching (NINT), a novel approach that uses the Neural Tangent Kernel to dynamically select training coordinates for faster convergence.

Findings

NINT nearly halves training time compared to existing methods.

NINT maintains or improves representation quality.

NINT achieves state-of-the-art acceleration among sampling strategies.

Abstract

Implicit Neural Representations (INRs) parameterize continuous signals via multilayer perceptrons (MLPs), enabling compact, resolution-independent modeling for tasks like image, audio, and 3D reconstruction. However, fitting high-resolution signals demands optimizing over millions of coordinates, incurring prohibitive computational costs. To address it, we propose NTK-Guided Implicit Neural Teaching (NINT), which accelerates training by dynamically selecting coordinates that maximize global functional updates. Leveraging the Neural Tangent Kernel (NTK), NINT scores examples by the norm of their NTK-augmented loss gradients, capturing both fitting errors and heterogeneous leverage (self-influence and cross-coordinate coupling). This dual consideration enables faster convergence compared to existing methods. Through extensive experiments, we demonstrate that NINT significantly reduces training time by nearly half while maintaining or improving representation quality, establishing state-of-the-art acceleration among recent sampling-based strategies.