Connectionist Modeling of Language: Examples and Implications

摘要: Researchers interested in human cognitive processes have long used computer simulations to try identify the principles of cognition. The strategy has been build computational models that embody putative and then examine how well such capture performance tasks. Until early 1980’s, this effort was undertaken largely within context “computer metaphor” mind. built based on conceptualization mind operated as though it were a conventional digital computer. However, with advent so-called connectionist, neural network, or parallel distributed processing (Anderson, Silverstein, Ritz, & Jones, 1977; Hinton Anderson, 1981; McClelland Rumelhart, McClelland, PDP Research Group, 1986; 1986), researchers began exploring implications are more broadly consistent style computation employed by brain. In connectionist models, take form cooperative competitive interactions among large numbers simple, neuron-like units (see Figure 1). Unit governed weighted connections collectively encode long-term knowledge system. activity some encodes input system; resulting other system’s response input. patterns remaining hidden constitute learned internal representations mediate between inputs outputs. Learning involves modifying values connection weights feedback from environment accuracy responses. While each unit exhibits non-linear spatial temporal summation, not generally considered be one-to-one correspondence actual neurons synapses. Rather, systems attempt essential properties vast ensembles real neuronal elements found brain, through smaller networks units. way, approach is distinct neuroscience (Sejnowski, Koch, Church