How random is the discharge pattern of cortical neurons? We examined recordingsfrom primary visual cortex (V1) and extrastriate cortex (MT) of awake,behaving macaque monkey, and compared them to analytical predictions. Wemeasured two indices of firing variability: the ratio of the variance to themean for the number of action potentials evoked by a constant stimulus, andthe rate-normalized Coefficient of Variation (C_v) of the interspike interval distribution.Firing in virtually all V1 and MT neurons was nearly consistentwith a completely random process (e.g., C_v ≈ 1).
We tried to model this high variability by small, independent, and random EPSPsconverging onto a leaky integrate-and-fire neuron (Knight, 1972). Boththis and related models predicted very low firing variability ( C_v ≪ 1) for realisticEPSP depolarizations and membrane time constants. We also simulateda biophysically very detailed compartmental model of an anatomically reconstructedand physiologically characterized layer V cat pyramidal cell with passivedendrites and active soma. If independent, excitatory synaptic input firedthe model cell at the high rates observed in monkey, the C_v and the variabilityin the number of spikes were both very low, in agreement with the integrate-and-fire models but in strong disagreement with the majority of our monkeydata. The simulated cell only produced highly variable firing when Hodgkin-Huxley-like currents (I_(Na) and very strong I_(DR) were placed on the distal basaldendrites. Now the simulated neuron acted more as a millisecond-resolutiondetector of dendritic spike coincidences than as a temporal integrator, therebyincreasing its bandwidth by an order of magnitude above traditional estimates.
This hypothetical submillisecond coincidence detection mainly uses the cell'scapacitive localization of very transient signals in thin dendrites. For millisecond-levelevents, different dendrites in the cell are electrically isolated from oneanother by dendritic capacitance, so that the cell can contain many independentcomputational units. This de-coupling occurs because charge takes timeto equilibrate inside the cell, and can occur even in the presence of longmembrane time constants.
Simple approximations using cellular parameters (e.g., R_m, C_m, R_i, G_(Na) etc)can predict many effects of dendritic spiking, as confirmed by detailed compartmentalsimulations of the reconstructed pyramidal cell. Such expressions allowthe extension of simulated results to untested parameter regimes. Coincidence-detectioncan occur by two methods: (1) Fast charge-equilization inside dendriticbranches creates submillisecond EPSPs in those dendrites, so that individualbranches can spike in response to coincidences among those fast EPSP's,(2) strong delayed-rectifier currents in dendrites allow the soma to fire onlyupon the submillisecond coincidence of two or more dendritic spikes. Such fastEPSPs and dendritic spikes produce somatic voltages consistent with intracellularobservations. A simple measure of coincidence-detection "effectiveness"shows that cells containing these hypothetical dendritic spikes are far moresensitive to coincident EPSPs than to temporally separated ones, and suggesta conceptual mechanism for fast, parallel, nonlinear computations inside singlecells.
If a simplified model neuron acts as a coincidence-detector of single pulses, networksof such neurons can solve a simple but important perceptual problem-the"binding problem" -more easily and flexibly than traditional neurons can.In a simple toy model, different classes of coincidence-detecting neurons respondto different aspects of simple visual stimuli, for example shape andmotion. The task of the population of neurons is to respond to multiple simultaneousstimuli while still identifying those neurons which respond to a particularstimulus. Because a coincidence-detecting neuron's output spike trainretains some very precise information about the timing of its input spikes, allneurons which respond the same stimulus will produce output spikes with anabove-random chance of coincidence, and hence will be easily distinguishedfrom neurons responding to a different stimulus. This scheme uses the traditionalaverage-rate code to represent each stimulus separately, while usingprecise single-spike times to multiplex information about the relation of differentaspects of the stimuli to each other: In this manner the model's highlyirregular spiking actually reflects information rather than noise.
