Knowledge-rich solutions to the binding problem: a simulation of some human computational mechanisms

摘要

The binding problem, how properties are represented as belonging to individuals, is identified as a severe problem for human memory, for which the memory adopts knowledge-rich solutions. It is argued that it is the nature of these solutions that endows human memory with many of its positive properties, particularly rapid retrieval on the basis of unreliable search clues. Parallel Distributed Processing (PDP) systems offer some insight into how human memory systems may work, as they also have to solve the binding problem by knowledge-rich methods. Experimental analysis and statistical models of Memory for Individuals Task (MIT) are presented, which provide evidence that the memory representations underlying human performance consist of sets of existential facts containing no referential terms. It is shown that the proposed representations can be incorporated directly into a PDP simulation of the inference from representation to response, and that the resulting system produces human-like errors when subjected to noisy input. The PDP simulation captures some of the asymmetries between stimulus and response which the statistical model cannot.