Many of the proposed models share the characteristic of decoding time using arrays of neural elements that differ in terms of some temporal property. The most generic of these is the spectral timing model of Grossberg and colleagues (Grossberg & Schmajuk 1989), which has been expressed in varying forms. The original model assumed a population of cells that react to a stimulus with an array of differently timed responses. Two variants of this motif have also appeared. One is a variant of clock models: Stimuli activate arrays of cells that oscillate at different frequencies and phases. By doing so, points in time following the onset of a stimulus can be encoded by activity in a subset of neurons that differs, at least somewhat, from the subsets of cells active at other times (Miall 1989, Gluck et al. 1990). In another model generally referred to as tapped delay lines, simple assumptions about connectivity lead to a sequential activation of different neurons at different times following a stimulus (Desmond&Moore 1988, Moore 1992, Moore&Choi 1997).

Spectral models have the advantage of encoding the time since the arrival of a stimulus by having different subsets of cells active at different times. Combined with simple learning rules where a teaching or error signal modifies connections for only active cells, spectral models can learn outputs that are properly timed and can even show the Weber effect of increased variance with increased delay.

However, we believe it is unlikely that spectral models are robust enough to generalize to complex temporal processing involved in speech and music recognition and complex motor patterns.