Di Bono, M. G. & Zorzi, M.
Department of General Psychology, University of Padova
Relative-position and transposition priming effects in visual word recognition have inspired alternative proposals about the nature of orthographic coding. The Open-Bigram model assumes that the relative position of a letter within a word is coded through its constituent ordered letter pairs. Alternatively, the Overlap model assumes that each letter is coded by a gaussian distribution of activation across the ordinal positions in a word. We asked what type of intermediate coding would emerge in a neural network learning location-invariant representations of written words. We trained a “deep” network with many layers on an artificial dataset of 120 words (trigrams) presented at five possible locations. The network learned a hierarchical generative model of the sensory input (unsupervised learning). We analysed the internal representations across layers as a function of input stimulus type (words, letters, bigrams). Word selectivity and location invariance increased as a function of layer depth. The activation pattern of each word was highly correlated with those of the first constituent letter and the constituent open bigram (i.e., the first and the last constituent letters). These results, though preliminary, suggest that bigram coding plays an important role in word recognition.