The varying severity of the patients’ impairments was simulated by altering the degree of damage. A neural network model was constructed and trained with the Light Efficient Network Simulator (Rohde, 1999). It was implemented as a simple-recurrent Elman

network model to reduce the computational demands (Plaut and Kello, 1999) and the exact computational architecture to realize this implementation is shown in Figure S1. Specifically, once a pattern was clamped to the sound input layer (in repetition/comprehension, for example), the activation spread to (1), iSMG → insular-motor layers; (2), to mSTG → aSTG → vATL layers; and (3), mSTG → aSTG → triangularis-opercularis → insular-motor layers at every time tick. The activation pattern at every layer was fed back to the previous layer Selleck Volasertib at the next time tick by utilizing the copy-back

connections to realize bidirectional connectivity (see Figure S1 for further details). A sigmoid activation function was used for each unit, bounding activation at 0 and 1. Eight hundred fifty-five high-frequency and eight hundred fifty-five low-frequency Japanese nouns, each three moras (subsyllabic spoken unit) in length, were selected from the NTT Japanese psycholinguistic database (Amano and Kondo, 1999)(see Supplemental Experimental Procedures, for the item properties). The remaining 3511 tri-mora nouns in the corpus were used for testing generalization. Each mora was converted into a vector of 21 bits representing pitch accent and Selleck BVD523 the distinctive phonetic-features

of its comprising consonant and vowel (the exact vector patterns are provided in Supplemental Experimental Procedures), following previous coding systems (Halle and Clements, 1983). Past simulations of language activities in English have used exactly the same coding scheme (Harm and Seidenberg, 2004) and so our findings should be language-general. The acoustic/motor representation of each word was made up of the three sequential, distributed mora patterns. Semantic representations were abstract vector patterns Methisazone of 50 bits generated from 40 prototype patterns, containing 20 “on” bits randomly dispersed across the 50 semantic units. Fifty exemplar patterns were generated from each of these prototypes by randomly turning off 10 of the 20 on bits, again following the previous coding systems of past English simulations (Plaut et al., 1996). Each semantic pattern was randomly assigned to one of the 1710 auditory patterns, ensuring an arbitrary mapping between the two types of representation. In repetition, each 21-bit mora vector was clamped to the input auditory layer, sequentially (i.e., one mora per tick), during which the insular-motor output layer was required to be silent. From the fourth tick (once the entire word had been presented), the output layer was trained to generate the same vector patterns sequentially (i.e., one mora per a tick), resulting in six time ticks in total.