Here we analyze natural variability within two types of systems. 1, The output of the biological spinal central pattern generator for locomotion in the cat, and 2, Sets of stochastic neural networks giving an output qualitatively similar to that observed within the biological system.

Fictive locomotion contains asymmetric transitions between the flexion and extension phases.  The transition from extension to flexion is: 1, Always strongly phase locked; 2, Composed of overlapping extensor burst offsets and flexor burst onsets; and 3, Invariant to changes in mean cycle period.  The transition from flexion to extension is:  1, Weakly phase locked within bouts containing short cycle periods, and well phase locked in bouts containing long cycle periods; 2, Offset times of flexor bursts and the onset times of extensor bursts do not overlap; and 3, Strength of phase locking depends critically upon relative timing of flexor offset and extensor onset.

Stochastic neural networks that qualitatively reproducing the timing relationships observed within the biological system have outputs that depend upon both the architecture of the network as well as model neuronal type (oscillatory-non-oscillatory).  Within models designed to reproduce the bi-phasic activity observed in some muscles, correlation of the bi-phasic burst is strongly influenced by model connectivity.  Additionally sets of leaky integrators have burst durations, which are sometimes well correlated even though they are well separated in time.  

Half-center models producing alternating output are strongly influenced by the internal structure of simulated neurons.  A half-center composed of a pair of leaky-integrators has transitions between phases which are always well phase locked, and overlapping.  Half-centers composed of intrinsically oscillatory Morris-Lecar neurons have transitions between phases whose phase locking is parameter dependent.  This parameter dependence is mainly due to changes in the timing of burst offset and burst onset.

We conclude that the output of the biological central pattern generator is likely to be strongly influenced by the intrinsically oscillatory properties of its neurons.  Models containing non-intrinsically oscillatory simulated neurons are unable to account for observed variability within the output of the biological system.