AI & ML Breaks Assumption

A systematic study reveals that SOTA representation learning methods for microscopy perform no better than untrained models or simple structural baselines.

arXiv · March 17, 2026 · 2603.13377

Ivan Svatko, Maxime Sanchez, Ihab Bendidi, Gilles Cottrell, Auguste Genovesio

The Takeaway

This challenges the assumption that 'biologically meaningful' features are being learned in bioimaging ML. It suggests that current benchmarks are masking a lack of progress and that the field needs a fundamental shift in how it evaluates visual representations for science.

From the abstract

Representation learning has driven major advances in natural image analysis by enabling models to acquire high-level semantic features. In microscopy imaging, however, it remains unclear what current representation learning methods actually learn. In this work, we conduct a systematic study of representation learning for the two most widely used and broadly available microscopy data types, representing critical scales in biology: cell culture and tissue imaging. To this end, we introduce a set o