FlyOCR’s creator reports 87% accuracy on sampled characters using a fruit fly brain model

FlyOCR is an optical character recognition experiment built on the MaleCNS v1.0 fruit fly connectome—a map of neural connections—according to its creator. They describe splitting PDF images into individual characters, mapping pixels to receptor activations and using a compact model to turn simulated neural spikes into text.

The creator reports 87% accuracy across 1,700-plus sampled characters and digits. In a test on a Microsoft 10-K filing, they say it scored about 86% on the balance sheet heading and read numeric values mostly correctly.