Roxana Ismail-Beigi¹, Yaoda Xu¹; ¹Yale University

A diverse array of writing systems has been developed across human culture and history. After prolonged training, humans are able to acquire reading proficiency in each of these writing systems. Here we asked to what extent distinguishing features in visual object perception are incorporated into the design of writing systems. Bao et al. (2020, Nature) showed that visual objects are differentiated primarily by their animacy/curvature and spikiness in the primate infero-temporal (IT) cortex. The same two shape-defining features are also present in convolution neural networks (CNNs) trained for object classification. Are writing systems designed to take advantage of these shape-defining features and maximize letter/character dissimilarities in curvature and spikiness to aid efficient letter learning and reading? Or are writing systems developed serendipitously without such considerations? To test this, we generated letter/character images from writing systems including alphabets (Latin, Greek, Cyrillic, Elder Futhark, Mongolian), abjads (Arabic, Hebrew, Phoenician), abugidas (Hindi Devanagari, Thai), and syllabaries (Japanese kana, Cherokee, Persian cuneiform), as well as a constructed abugida (Tengwar). We ran these images through the ImageNet-trained Alexnet layer FC6 and plotted the resulting activations in an object space defined by curvature and spikness. This object space was constructed following Bao et al. and closely resembled the one obtained from the macaque IT cortex neuronal responses. We then obtained the average pairwise distances of the letters/characters for a given writing system in this space. Among the writing systems examined, Latin letters exhibit the greatest dissimilarity in terms of curvature and spikiness, with their overall distribution roughly expanding the object space. Other writing systems, such as kana, Devanagari, and Thai, tend to cluster densely in one area of the object space. Thus, while some writing systems incorporate the defining features of object perception, others appear to do so to a lesser extent.

Acknowledgements: NIH grant R01EY030854