Conference on Cognitive Computational Neuroscience

Learning to read fluently is associated with the precise encoding of letter position that enables us to distinguish anagrams such as TRIAL and TRAIL. However, the neural basis of how letters and their positions are encoded in our brain is yet unknown. In this study, we trained deep neural network-based models on different languages and found that some units in the later layers of the network are tuned to letter identity. Next, by dissociating the absolute word position, with the relative position of a preferred letter within a word, we tracked how these networks transit from retinotopic to ordinal letter-position coding scheme. Our simulations posit that the units in early layers encode blank spaces at the edges of a word, which are pooled together to encode the ordinal positions of the edge letters irrespective of word position. These predictions are further validated using fMRI studies. Overall, we propose a new schema on how letter positions might be encoded in our brain.