![The spectral neuron - an ML primitive for scalable and interpretable models [R]](https://external-preview.redd.it/5FeXCLfMGjVAviSsjMffsXKjmE4cHIxniVezsEMG_tU.png?width=960&crop=smart&auto=webp&s=26f5c6439b8aa474e757991880179a326d2b9377)
The spectral neuron - an ML primitive for scalable and interpretable models [R]
Worked some time ago on one of the ad teams at Yahoo, and this grew out of a question I kept returning to while there are there "simple" models that are both simple, scalable, interpretable, and controllable at the same time?
Decided to explore it, first in a blog (starting here), then in a new preprint "The Spectral Neuron", built by distilling latest blog-posts into a manuscript, I study models of the form:
𝑓(𝒙) = 𝛌ₖ(𝐀₀ + 𝚺ᵢ 𝑥ᵢ𝐀ᵢ).
Manuscript: https://arxiv.org/abs/2608.08003
Code: https://github.com/alexshtf/spectral_neuron_paper
Looks like a simple on-liner, but many interesting aspects hide there. How expressive does the model become as the matrices grow? What can we read directly from the learned matrices? Which shapes can be guaranteed by construction?
I develop the mathematics, give a practical initialization and training recipe, and test the model in scaling experiments on synthetic and real data.
AI disclaimer: manuscript written by yours truly, AI assisted in looking up canonical references and related work for literature review. In contrast, the code was heavily AI written and reviewed by yours truly.