# Emergent Denoising of SDSS Galaxy Spectra Through Unsupervised Deep Learning

**Authors:** Oliver Camilleri, Zahra Sharbaf, Ignacio Ferreras  
**arXiv:** [2510.08411](https://arxiv.org/abs/2510.08411)  
**Date:** 2025-10-09  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2510.08411v1/

## Key points

- Authors discuss challenges in analyzing absorption spectra in high redshift or low mass galaxies due to low signal-to-noise ratios
- Proposal to use unsupervised Deep Learning techniques to denoise galaxy spectra
- Training methods on diverse sample of spectra with varying signal-to-noise levels to reconstruct higher quality spectra
- Study focuses on three key line strengths and compares results with synthetic data to assess retrieval biases
- Standard Autoencoder identified as powerful method for denoising galaxy spectra without introducing systematic errors
- Emphasis on careful analysis to avoid overfitting noisy input data and biased reconstructions
- Research highlights potential of Deep Learning in enhancing understanding of galaxy evolution through improved spectral analysis techniques
- Contribution of valuable insights into optimizing observational methods for studying star formation and chemical enrichment histories across different redshifts and mass ranges.

## Summary

, , , , 

In their paper titled "Emergent Denoising of SDSS Galaxy Spectra Through Unsupervised Deep Learning," authors Oliver Camilleri, Zahra Sharbaf, and Ignacio Ferreras discuss the challenges associated with extracting information from galaxy spectra. They focus on the difficulties in analyzing absorption spectra in high redshift or low mass galaxies due to low signal-to-noise ratios. This can make accurate estimates of stellar population parameters challenging for traditional spectroscopic surveys. To address this issue, the authors propose using unsupervised Deep Learning techniques to denoise galaxy spectra. By training these methods on a diverse sample of spectra with varying signal-to-noise levels, they are able to reconstruct higher quality spectra and explore the potential of Deep Learning in accurately reproducing data from incomplete sources. The study specifically focuses on three key line strengths and compares the results with synthetic data to assess any retrieval biases. The findings suggest that a standard Autoencoder is a powerful method for denoising galaxy spectra without introducing systematic errors in the reconstruction process. However, the authors emphasize the importance of careful analysis as some methods may produce seemingly noiseless spectra that are actually biased due to overfitting noisy input data. This research highlights the potential of Deep Learning in enhancing our understanding of galaxy evolution through improved spectral analysis techniques. It also contributes valuable insights into optimizing observational methods for studying star formation and chemical enrichment histories in galaxies across different redshifts and mass ranges. By employing denoising techniques with minimal bias, ongoing and future spectral surveys such as DESI, WEAVE, and WAVES can maximize their quality and accuracy.