ABSTRACT
To what extent can information distributed across the visual field be processed in parallel? A connectionist model, capable of recognising multiple words appearing simultaneously on its “retina,” is described which addresses this question. The model relies on the notion of a hierarchy of detectors, starting at the lowest level with position-specific primitive-feature detectors, and progressing to a level composed of position-independent “letter cluster” detectors. Intervening levels register successively higher-order features and also collapse over local spatial regions of the level below, resulting in less positional specificity of the detectors. Using an associative learning rule, the model has been taught to recognise a large sample of words in arbitrary retinal locations. Following this training, it is also able to recognise several words simultaneously, although under certain conditions crosstalk among words can become unmanageable. The model includes an attentional mechanism, which can limit crosstalk, and a serial readout mechanism, which is necessary for a word to reach awareness. While exhaustive simulation experiments have yet to be carried out, there are a variety of phenomena, both experimental and anecdotal, that the model appears well-equipped to account for, including: translation and scale invariant recognition, positional uncertainty at the letter and word levels, the recognition of misspelled words, the integration of information across fixations, similarity-based interference effects, and the role of focal attention in localisation.
