EVOS: Efficient Implicit Neural Training via EVOlutionary Selector

TL;DR

EVOS introduces an efficient training paradigm for Implicit Neural Representations by evolving coordinate subsets used in forward passes. Through sparse fitness evaluation, frequency-guided crossover, and augmented unbiased mutation, EVOS reduces training time by up to 48-66% without sacrificing reconstruction quality, outperforming existing sampling-based acceleration methods. The approach delivers broad compatibility across backbones and network sizes, and ablation studies confirm the necessity of each component. By reframing sampling as a dynamic evolutionary process, EVOS not only lowers computation but also provides competitive or superior signal fitting performance across 1D, 2D, and 3D tasks, with practical implications for faster INR-based applications.

Abstract

We propose EVOlutionary Selector (EVOS), an efficient training paradigm for accelerating Implicit Neural Representation (INR). Unlike conventional INR training that feeds all samples through the neural network in each iteration, our approach restricts training to strategically selected points, reducing computational overhead by eliminating redundant forward passes. Specifically, we treat each sample as an individual in an evolutionary process, where only those fittest ones survive and merit inclusion in training, adaptively evolving with the neural network dynamics. While this is conceptually similar to Evolutionary Algorithms, their distinct objectives (selection for acceleration vs. iterative solution optimization) require a fundamental redefinition of evolutionary mechanisms for our context. In response, we design sparse fitness evaluation, frequency-guided crossover, and augmented unbiased mutation to comprise EVOS. These components respectively guide sample selection with reduced computational cost, enhance performance through frequency-domain balance, and mitigate selection bias from cached evaluation. Extensive experiments demonstrate that our method achieves approximately 48%-66% reduction in training time while ensuring superior convergence without additional cost, establishing state-of-the-art acceleration among recent sampling-based strategies.

Figures (10)

• Figure 1: Overview of EVOS framework. The proposed method aims to optimize a MLP for implicit signal representation via evolutionary selector. The process comprises three key components: (1) (Sparse) Fitness Evaluation (Sec. \ref{['fitness']}) for efficiently guiding coordinate selection, (2) (Frequency-Guided) Crossover (Sec. \ref{['crossover']}) for improving performance by balancing frequency domain preferences, and (3) (Augmented Unbiased) Mutation (Sec. \ref{['mutation']}) for mitigating selection bias in each iteration. The selected coordinates from this evolutionary process are then fed into the network, enabling sparsified forward passes and reduced computational costs.
• Figure 2: Visualization of Constant (Top) and Step-wise (Bottom) schedulers. Constant maintains fixed selection ratio throughout iterations, while Step-wise implements progressive increase.
• Figure 3: Visual comparison of sampling-based acceleration methods for INR training. All methods are evaluated under identical conditions with a fixed training duration of 60 seconds.
• Figure 4: Comparison of reconstruction quality (PSNR & SSIM) across different sampling-based methods over training time.
• Figure 5: Visualization of performance improvements across different backbone architectures when integrated EVOS. All experiments are conducted under consistent conditions with a fixed training duration of 120 seconds.