CORread

Train models of the ventral visual stream to read, and study the word-selective units that emerge.

Learning to read repurposes part of the brain's object recognition system: a patch of ventral occipito-temporal cortex, the visual word form area, becomes selective for written words. Hannagan et al. (2021) reproduced this "neuronal recycling" in CORnet-Z, a convolutional network of the ventral stream. Trained first on images and then on images and words together, the network develops a small set of word-selective units with a compositional code for letters and their positions.

CORread turns that experiment into a reproducible, extensible library:

• Stimuli: word images, and word sets that can be regenerated byte for byte from a recipe.
• Training: pre-literate, illiterate and literate CORnet-Z and CORnet-S networks, resumable, each run fully recorded.
• Lesions: word selectivity of the dense-layer units, lesions of them, and the recovery of reading.
• Sharing: models and word set recipes through a Hugging Face organisation of your own.
• Reporting: the API contract for following runs live on a dashboard of your own.

Start with Getting started. The API reference documents every public class and function.

CORread is free software under the GNU GPL v3 or later. It derives from CORnet (DiCarlo Lab) and Origins-of-VWFA (Hannagan et al.), via Dr. Olga Solaja's adaptation; see the README for credits and citations.