Carnegie-Mellon University

• Main Authors: Geoffrey E. Hinton, Terrence J. Sejnowski
• Other Authors: The Pennsylvania State University CiteSeerX Archives
• Format: Text
• Language: English
• Subjects: Ballard; The Slot; The Pointers
• Online Access: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.164.1437; http://papers.cnl.salk.edu/PDFs/Learning%20Semantic%20Features%201984-3235.pdf

An important idea within cognitive science is that much general knowledge can be represented as constraints between the slot-fillers of a schcma. The central idea of "connectionism " is that knowledge is represented by the strengths of the connections in a large nctwork of simple processing elements. The relation between these two ideas is complex. Several ways of using connectionist networks to implement schemas have been proposed. The obvious, "localist " approach is to identify processing units in the physical network with concepts, and to treat the physical links as if they were direct implementations of the pointers that are conventionally used to represent the filling of a schema-slot by an object (Feldman and Ballard, 1982; Fahlman, 1979). An alternative, "distributed approach is to allocate a large number of units to each slot of a schema, and to represent the filler of that slot by the pattern of activity of that set of units (Hinton, 1981). The main difference is in how the physical parallelism is used. In the distributed approach, only one instantiation of a particular schema is possible at a time because the units dedicated to each slot can only have one pattern of activity at a time. The physical links are used to implement constraints between slot-fillers. By setting the strengths of the links appropriately, it is possible to make a pattern of activity in one set of units cause (or prohibit) a pattern in another set of units. If each component of a pattern of activity is viewed as a semantic feature of the object