OmniSpectra encodes spectra by wavelength instead of token rank
Original title: OmniSpectra: A Unified Foundation Model for Universal Spectra Representation Learning
The 12.2M-parameter model, pretrained on eight surveys, beats seven baselines and larger models.
OmniSpectra is a wavelength-native foundation model for spectral representation learning. It uses a continuous-coordinate masked autoencoder so spectra from different instruments are not resampled onto integer token positions. A logarithmic coordinate basis turns cosmological redshift into an additive shift that cancels in continuous rotary embeddings, and a resolution-gated Fourier decoder is meant to limit spectral bias. Pretrained on eight surveys, the 12.2M-parameter model is reported to reach state-of-the-art results against seven baselines while beating substantially larger models.
- The paper treats integer pixel indexing as a cross-instrument limitation.
- Log coordinates convert multiplicative redshift scaling into a uniform additive shift.
- Authors say physical structure in the encoding can substitute for parameter scale.
- AI models
- OmniSpectra
From the original article
1 min read at huggingface.co
Standard sequence models encode position by token rank rather than physical coordinate, typically forcing practitioners to destructively resample continuous physical signals such as spectra observed across heterogeneous instruments.
By Md Khairul Islam, Zeyu Xia, Adam S. Bolton · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.