Abstract

The Visual Word Form Zone (VWFA) is a region of the human inferotemporal cortex that emerges at a fixed location in the occipitotemporal cortex during reading acquisition and consistently responds to written words in literate individuals. According to the neural recycling hypothesis, this region results from the reassignment, for letter recognition, of a subpart of the ventral visual pathway initially involved in the recognition of faces and objects. Moreover, according to the biased connectivity hypothesis, its reproducible localization is due to pre-existing connections of this sub-region to the areas involved in the processing of spoken language. Here, we evaluate these assumptions in an explicit computational model. We trained a deep convolutional neural network of the ventral visual pathway, first to categorize images, and then to recognize words written invariantly for case, font, and size. We show that the model can take into account many properties of VWFA, especially when a subset of units has biased connectivity to word output units. The network develops a sparse and invariant representation of written words, based on a restricted set of selective reading units. Their activation mimics several properties of VWFA, and their lesion causes a specific reading deficit. The model predicts that, in literate brains, written words are encoded by a compositional neural code with neurons tuned to either individual letters and their ordinal position relative to the start or end of the word or to pairs of letters. (bigrams).