[논문 리뷰] Robustness of Neural Networks for CMB Polarization Foreground Removal

Luca Gomez Bachar, Cora Dvorkin외 1명|arXiv (Cornell University)|2026. 03. 12.

Cosmology and Gravitation Theories인용 수 0

한 줄 요약

논문은 CNN 기반 전경 제거가 CMB 편파(CMB polarization)에서 서로 다른 Galactic 전경 모델(FM)들에 걸쳐 일반화되는 방식을 분석하고, 더 복잡한 FM에서 학습하는 것이 보지 못한 모델에 대한 견고성을 향상시킨다는 것을 보여준다.

ABSTRACT

The detection of Cosmic Microwave Background primordial $B$-mode polarization would constitute a ``smoking gun" signal of primordial gravitational waves. However, this measurement requires accurate removal of polarized Galactic foregrounds to avoid systematic biases when estimating the tensor-to-scalar ratio. Methods based on Machine Learning techniques (ML), such as Convolutional Neural Networks (CNNs), have recently been proposed as alternative foreground cleaning techniques, but their applicability to real data relies on their ability to generalize beyond the models assumed during training. In this work, we focus on a variety of foreground models (FMs) used for training and conduct a systematic study of the generalization properties of a CNN-based method. We train various CNN architectures on simulations generated from different Galactic FMs, and test their performance on models not used during the training. By characterizing the statistical properties of the FMs using variance, skewness, and Shannon entropy, we define a statistical complexity hierarchy among them. We show that training on the more complex FMs reduces bias and improves precision when testing on unseen FMs, whereas training on the simplest model could introduce systematic errors. These results evidence that a lack of generalization is a relevant source of systematic uncertainty, and emphasize the importance of understanding the impact of the models assumed during training in ML-based methods before applying them to real data.

연구 동기 및 목표

• Assess the generalization of CNN-based foreground removal to FM(s) not seen during training.
• Quantify how foreground model complexity affects CMB reconstruction accuracy and precision.
• Define statistical metrics to characterize foreground models and link them to ML generalization performance.

제안 방법

• Generate lensed CMB Q/U maps, instrumental noise, and Galactic foregrounds using multiple foreground models.
• Train CNN architectures (UN, UB, L3) on simulations from specific FMs and test on unseen FMs.
• Preprocess Healpy maps into CNN-friendly 2D blocks and optimize using MAE plus FFT-based physics term in the loss.
• Evaluate performance with Cℓ-based ratios between reconstructed and true CMB maps across realizations.
• Characterize FM complexity via variance, skewness, and Shannon entropy to interpret generalization behavior.

실험 결과

연구 질문

• RQ1 How does CNN foreground cleaning generalize when trained on one FM and tested on a different FM?
• RQ2 Does increasing foreground model complexity during training improve reconstruction accuracy and reduce bias on unseen models?
• RQ3 What FM statistics (variance, skewness, entropy) correlate with CNN generalization performance?
• RQ4 Can CNNs utilize high-resolution frequency data without degrading angular resolution like traditional methods?
• RQ5 How do different CNN architectures compare in terms of stability and performance for this task?

주요 결과

• Training on more complex foreground models reduces bias in the reconstructed CMB when tested on unseen FMs.
• Training on the simplest FM can introduce systematic errors in CMB reconstruction across unseen models.
• The study identifies RP as the most statistically complex FM, while GP and d11s6 show similar complexity, impacting generalization.
• Pixel-level FM statistics (variance, skewness, Shannon entropy) correlate with generalization performance and help interpret results.
• CNN-based foreground cleaning preserves high-resolution information by avoiding forced resolution degradation used in some traditional methods.