Beyond Uniform Sampling: Synergistic Active Learning and Input Denoising for Robust Neural Operators

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Engineering & Applied Sciences · Depth: Expert, extended

Summary

Researchers from the University of Illinois Urbana-Champaign and the Indian Institute of Technology Delhi have developed a synergistic defense combining active learning and input denoising to enhance the adversarial robustness of neural operators, which are critical for safety-critical digital twin deployments. Neural operators, while fast surrogate models for physics simulations, are highly vulnerable to adversarial perturbations, leading to catastrophic deviations from true physics. The proposed method, tested on the viscous Burgers' equation benchmark, integrates an active learning component that adaptively probes model weaknesses using differential evolution attacks and generates targeted training data with physics-corrected labels. It also includes an input denoising architecture that uses a learnable autoencoder bottleneck to filter adversarial noise. This combined approach achieved a 2.04% combined error (1.21% baseline + 0.83% robustness), representing an 87% reduction compared to standard training's 15.42% combined error, significantly outperforming either active learning alone (3.42%) or input denoising alone (5.22%).

Key takeaway

For Machine Learning Engineers developing neural operators for safety-critical applications like nuclear digital twins, relying solely on standard training or validation accuracy is insufficient. You should integrate both active learning for targeted data generation and architectural input denoising to achieve robust models that accurately track physics under perturbed conditions, significantly reducing error and enhancing reliability in deployment.

Key insights

Combining active learning with input denoising synergistically improves neural operator robustness against adversarial attacks.

Principles

• Optimal training data for neural operators is architecture-dependent.
• Robustness evaluation must be explicit, not inferred from validation accuracy.

Method

An iterative active learning loop probes model weaknesses via differential evolution, generates targeted training data with physics-corrected labels, and uses an adaptive smooth-ratio safeguard. An input denoising layer with a learnable autoencoder bottleneck filters high-frequency adversarial perturbations.

In practice

• Use physics-corrected labels for adversarial training.
• Implement learnable blend weights in denoising architectures.
• Target training data generation to discovered model vulnerabilities.

Topics

• Neural Operators
• Adversarial Robustness
• Active Learning
• Input Denoising
• Digital Twins

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Beyond Uniform Sampling: Synergistic Active Learning and Input Denoising for Robust Neural Operators — Combining active learning with input denoising significantly enhances neural operator robustness against adversarial attacks.
• Are Candidate Models Really Needed for Active Learning? — Randomly initialized deep learning models can effectively select informative samples for active learning without candidate models.
• Deep Active Re-Labeling: Toward Noise-Resilient Annotation Efficiency — Deep Active Re-Labeling improves active learning by strategically re-annotating noisy data to enhance efficiency and data quality.
• Stability and Discretization Error of State Space Model Neural Operators — Theoretical bounds quantify discretization error and stability in neural operators for solving PDEs.
• Twincher: Bijective Representation Learning for Robust Inversion of Continuous Systems — Twincher enables robust inversion of continuous systems by learning bijective, perturbation-insensitive representations.
• A Mutual Information Lower Bound for Multimodal Regression Active Learning — A Two-Index framework and MI-LB acquisition function improve active learning for multimodal continuous regression.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.