Persistent neural activity is a sustained change in sodium
action potential firing that has been observed in many
brain areas involved in short term memory. We are studying
the oculomotor neural integrator where persistent activity
is a neural correlate of the short-term memory of eye
position. The experimental preparation is the goldfish,
which is particularly advantageous for a cellular and
computational analysis of mechanisms. We find no evidence
for plateau potentials or intrinsic oscillatory dynamics,
two hypothesized cellular mechanisms for persistent neural
activity. Partial pharmacological inactivation and analysis
of changes in the rate of synaptic potentials are consistent
with network mechanisms based on recurrent synaptic excitation.
Finally, a visual training stimulus can profoundly alter
the dynamics of persistent neural activity. The dependence
on stimulus parameters suggests visual input is normally
used to tune-up the neural integrator for better performance.
To test the reverberation hypothesis for persistent activity,
the oculomotor integrator has been modeled as a network
of conductance-based neurons interacting by recurrent
excitation. The strength of synaptic feedback is tuned
so that the network realizes "linear attractors" where
the level of persistent activity varies linearly with
the gaze position. The precision with which synaptic weights
need to be fine-tuned depends on the time constant of
the recurrent synapses, as it was emphasized previously
by Seung. The issue of robustness for continuous attractors
is discussed. These insights from computational modeling
may be applied to short-term memory networks for a continuous
quantity (in the form of continuous attractors) in general.
