Bottom-up Connectionist Models of 'Interaction': Dennis Norris
Although connectionism is neutral with respect to interaction, most early connectionist models of language processing were interactive. Theories such as Rumelhart and McClelland's (1982) model of visual word recognition, McClelland and Elman's (1986) TRACE model of speech recognition, and Dell's (1985; 1986) theory of speech production were all interactive-activation models. The interactive-activation models are hardwired networks with no ability to learn. Most new connectionist models incorporate a learning mechanism. The current modal model in connectionism is the three layer back-propagation network. Although it is possible to construct interactive models using back-propagation, most models use simple feed-forward (bottom-up) networks with a single layer of hidden units. Even a relatively complex model such as the Seidenberg and McClelland (1989) model of reading aloud has this simple structure.