Running Head: MODELS OF ORTHOGRAPHIC PROCESSING 1 Computational Models of Location-Invariant Orthographic Processing

We trained three topologies of backpropagation neural networks to discriminate 2000 words (lexical representations) presented at different positions of a horizontal letter array. The first topology (zero-deck) contains no hidden layer, the second (one-deck) has a single hidden layer, and for the last topology (two-deck), the task is divided in two subtasks implemented as two stacked neural networks, with explicit word-centered letters as intermediate representations. All topologies successfully simulated two key benchmark phenomena observed in skilled human reading: transposed-letter priming and relative-position priming. However, the two-deck topology most accurately simulated the ability to discriminate words from nonwords, while containing the fewest connection weights. We analyzed the internal representations after training. Zero-deck networks implement a letter-based scheme with a position bias to differentiate anagrams. One-deck networks implement a holographic overlap coding in which representations are essentially letter-based and words are linear combinations of letters. Two-deck networks also implement holographic-coding. 1. Introduction As they read text, skilled readers of languages that use an alphabetic script must map retinal images of letters onto abstract word representations. More specifically, learning to read involves the recognition of co-occurring letters as part of larger entities, that is, words. Reading also involves an ability to recognize words (strings of co-occurring letters) at various locations on the retina 1 , thus achieving location-invariant word recognition. Among the proposed models of the complex processing involved in this cognitive task, many posit a hierarchical system of increasing invariance, in which simple visual features are gradually integrated into more abstract and complex features (e.g., Dehaene, Cohen, Sigman, & Vinckier, 2005). For instance, visual features can be combined into representations of letters which are dependent on their physical attributes (e.g., font, case and location). Then, gradual abstraction from these physical attributes is achieved. High in the hierarchy, abstract (i.e., shape-invariant) letter representations are combined in a word-centered position coding scheme (see below) to finally activate location-invariant lexical representations, that is, words.