ABSTRACT
The problem of how neural systems represent and process temporal patterns such as language and music is central to cognition (Port 1990; Elman 1990; Jones 1976). Although a number of solutions have been proposed, this problem is far from being solved. The most common approach is to represent time by transforming it into a spatial dimension (Elman and Zipser, 1988). In this approach, temporal patterns are processed in chunks by collecting input for a fixed number of time steps and placing this information in a ‘buffer’, the size of which is fixed a priori. Time is represented explicitly by the input’s serial position within the buffer; i.e., the input at time t is adjacent to the input at time t + 1. There are a number of problems with this approach which have been elaborated on by Port (1990) and Elman (1990). A system using a spatial representation requires an interface with the world which is able to correctly guess the appropriate buffer size for each input pattern. An incorrect guess may result in relevant information dropping off the end of the buffer. Another problem with this approach is that it is inflexible to changes in pattern presentation rate. Patterns that are faster or slower than the ‘training’ rate have different input representations and hence have different effects on processing. This does not seem to be the case with language as we are capable of adapting our perceptual systems to different language production rates.
