UnEZ: How an Unsupervised Neural Network 'Grows' Redshift from Spectra

Traditional spectral redshift estimation has long relied on template spectra or manually curated catalogs, and both approaches have hit bottlenecks in the era of large-scale surveys. A new paper proposes the UnEZ model, which takes a completely different path: without relying on any redshift or line-intensity labels, it lets redshift naturally "emerge" from the spectral data.

When Neither Templates Nor Catalogs Are Enough

As spectroscopic surveys continue to scale up, the computational burden of estimating properties such as galaxy metallicity, star formation rate, and redshift keeps growing. Existing methods either rely on template spectra for matching or on curated redshift catalogs for supervised training. The problem is that both types of prior knowledge struggle to cover the diversity of objects appearing in large-scale surveys—templates are limited, and catalogs are biased.

UnEZ's core claim is that if the network architecture itself embeds sufficient physical constraints from spectroscopy, redshift and line intensities need not be fed to the model as labels—they can be learned from the structure of the spectra themselves.

Spectral redshift is an astronomical method for inferring a galaxy's recessional velocity by measuring the wavelength shift of known spectral lines (such as hydrogen Balmer lines, oxygen emission lines, etc.) in a galaxy's spectrum. According to the Doppler effect, light emitted by an object moving away from the observer has its wavelength stretched; the redshift z is defined as the ratio of observed wavelength to rest wavelength minus one. Redshift is a key observable for measuring cosmic expansion, galaxy distances, and the large-scale structure of the universe. Traditionally, redshift estimation relies on template matching: cross-correlating observed spectra with a set of template spectra at known redshifts to find the best match; or using supervised learning, training models with existing redshift catalogs as labels. These methods perform well when spectral quality is high and object types are familiar, but when facing the large numbers of faint, low-SNR, or rare-type objects in large surveys, insufficient template coverage and catalog selection bias become major sources of systematic error.

Encoder-Decoder Architecture: Predicting Lines, Not Redshift

UnEZ's design is worth a closer look. It consists of an encoder and a fixed deterministic decoder. The encoder predicts the amplitudes and widths of emission and absorption lines, and the decoder uses these line parameters to construct a uniquely corresponding template, estimate redshift, produce a redshift confidence curve, and generate a spectral model.

The key here is "self-templating": the model does not load an external template library but instead generates templates on the fly from its own predicted line parameters. Redshift is not a directly regressed target but a byproduct of the decoding process. This design allows the model to output interpretable intermediate quantities—line intensities and redshift confidence—even without any redshift labels.

Emission and absorption lines are the most important features in galaxy spectra. Emission lines arise from stimulated emission by gas—such as Hα, Hβ, [O III], and [N II]—and typically reflect star formation activity or active galactic nuclei; absorption lines form when atoms in stellar atmospheres or the interstellar medium absorb light at specific wavelengths, such as Ca II H&K, Na I D, and Mg Ib. Each line can be described by its amplitude (intensity), width (velocity dispersion), and central wavelength. In redshift estimation, the ratio of a line's observed wavelength to its rest wavelength directly gives the redshift value, so accurately measuring line parameters is equivalent to indirectly obtaining redshift information. UnEZ has the encoder predict these line parameters rather than directly regressing redshift, precisely exploiting the physical correspondence between lines and redshift, which makes the model's outputs interpretable.

Cross-Validation on SDSS DR18

The research team chose the Sloan Digital Sky Survey (SDSS) Data Release 18 as the testbed because the dataset is large and its redshift catalog is thoroughly validated, making it suitable as an independent benchmark for evaluating the model. This is methodologically important: UnEZ does not use redshift labels during training, and the SDSS catalog is used only for independent post-hoc comparison.

At signal-to-noise ratio SNR > 10, UnEZ's outlier fraction is 0.459% and its normalized median absolute deviation (NMAD) is 1.237 × 10⁻⁴. At the lower SNR > 3 condition, the outlier fraction is 3.89% and NMAD is 1.471 × 10⁻⁴. Considering that the model uses no templates, no line-intensity labels, and no redshift labels at all, this performance is competitive with traditional redshift estimation algorithms.

Outlier fraction and normalized median absolute deviation (NMAD) are two standard metrics for assessing the accuracy of photometric or spectroscopic redshift estimates. Outlier fraction is typically defined as the proportion of samples whose redshift estimation error exceeds a threshold (such as |Δz| > 0.15 or |Δz|/(1+z) > 0.15), reflecting the frequency of severe misestimates. NMAD measures the dispersion of the error distribution and is insensitive to outliers; it is computed as 1.4826 times the median absolute deviation, with smaller values indicating more precise estimates. In the astronomical literature, an NMAD on the order of 10⁻⁴ is generally considered spectroscopic-level precision, while values on the order of 10⁻³ are common for higher-quality photometric redshifts. UnEZ's NMAD of 1.237×10⁻⁴ at SNR>10 indicates precision approaching spectroscopic redshift levels, entirely without redshift labels.

The Spillover Value of an Unsupervised Model

The paper also demonstrates other uses of UnEZ's outputs. Because the model can predict line intensities and redshift, these intermediate results can in turn be used to construct template spectra needed by traditional redshift estimation algorithms. In other words, an unsupervised model can serve as a template generator, providing inputs to existing template-dependent pipelines.

The broader significance of this approach lies in scalability. Future surveys will produce enormous numbers of spectra, and the cost of manually maintaining template libraries and catalogs will keep rising. UnEZ demonstrates a template-agnostic, label-free path: redshift and other galaxy properties can emerge naturally from the spectra themselves, without having to presuppose a template space.

Self-templating is a mechanism that lets a model generate the reference templates it needs, as opposed to the traditional practice of loading fixed templates from an external template library. In UnEZ, the line parameters (amplitude and width) predicted by the encoder are fed into a deterministic decoder, which uses these parameters to construct on the fly a template uniquely corresponding to that spectrum. Because the template is generated by the model itself and does not depend on external priors, the entire pipeline is insensitive to the coverage and biases of template libraries. This design also gives the model's intermediate outputs—line intensities and the redshift confidence curve—physical meaning, allowing them to be directly reused by other algorithms. The self-templating idea has generalization value in unsupervised spectral analysis, especially for handling the large numbers of unknown-type objects in future surveys.

Questions Worth Watching

Being fully unsupervised means there is no "anchor" provided by labels, so how to guarantee the uniqueness and consistency of the model's physical scale is a question such methods must continually answer. The paper uses the SDSS catalog as an independent benchmark, which somewhat alleviates this concern, but consistency across surveys and instruments still awaits further validation.

Another point of interest is the practical usability of the redshift confidence curve—if the confidence can reliably filter out low-quality estimates, its value in pipelines would be significantly higher than outlier-fraction metrics alone. This may be a direction worth exploring in future work.

One-line Summary

UnEZ uses an unsupervised encoder-decoder to let redshift emerge from spectra, achieving competitive accuracy on SDSS.

Paragraph Summary

UnEZ is a fully unsupervised, physics-driven spectral redshift estimation model that uses no redshift labels, line-intensity labels, or external template libraries during training. Its encoder predicts the amplitudes and widths of emission and absorption lines, and a fixed deterministic decoder uses these to generate a unique template and derive redshift, a confidence curve, and a spectral model, realizing a self-templating mechanism. On SDSS DR18, at SNR>10 the outlier fraction is 0.459% and NMAD is 1.237×10⁻⁴; at SNR>3 the outlier fraction is 3.89% and NMAD is 1.471×10⁻⁴, competitive with traditional algorithms. The model's outputs can also feed back into traditional pipelines to construct template spectra, demonstrating the scalability of a template-agnostic, label-free path.