Shared geometry in astronomy models does not mean shared physics
Shared Geometry Is Not Shared Physics: A Layerwise Test of the Platonic Representation Hypothesis in Astronomy
Machine Learning
Summary
This paper looks at whether different AI models trained on space images and data understand the physical properties of objects in similar ways. The authors found that while the models often align in how they represent shapes or structures, this doesn’t mean they agree on the underlying physics. Different layers of the models show this mismatch, and models that align geometrically do not perform as well when predicting physical characteristics. This means that similar shapes detected by models do not guarantee a shared scientific understanding.
What this means in practice
- •For astronomical data engineers: Evaluate and select model layers based on label performance rather than geometric alignment for better physical property prediction in multi-survey data integration.
- •For machine learning engineers: Design transfer learning strategies acknowledging that geometric feature alignment does not imply consistent physical property encoding across models.
Authors
Kshitij Duraphe, Aravind Kannappan, Dun Li Chan, Okiki Famutimi, Yaswant Sai Ejjagiri, Michael J. Smith, John F. Wu
Abstract
In this paper we investigate whether geometrically aligned astronomical representations are also scientifically interchangeable. We analyze every layer's representation from 36 pretrained models using cross-matched optical images, infrared images, and spectra. All 139 analyzed model-survey comparisons show significant local-neighborhood alignment somewhere in the network after accounting for the search over depth. However, alignment does not generally increase with network depth and can be stronger for the changes between consecutive computational blocks than for the block outputs themselves. We find that geometry-selected stages under-perform label-selected stages in every image-transfer comparison; moreover, every final HSC-COSMOS-Web model pair is aligned while every redshift transfer has negative R^2. Models can therefore recover a similar cross-survey geometry among astronomical objects without recovering a survey-invariant linear encoding of the physical properties tested here.