We asked what the recurrent PING loop does before introducing training, classification or a learned readout. We compared the same Poisson-driven excitatory–inhibitory architecture with reciprocal feedback disabled and enabled.
Enabling the loop produced rhythmic population activity and strongly suppressed excitatory firing across the matched-drive sweep. This establishes the circuit’s basic operating behaviour, but the limited simulations do not show that gamma timing itself caused the suppression.
These cellular traces illustrate reciprocal E→I→E feedback; they do not measure population-wide phase locking or spikes per cycle.
We compared untrained loop-off and loop-on networks using the following simulation and measurement procedure.
Construct the two loop conditions. Both networks contained 1024 excitatory (E) and 256 inhibitory (I) neurons. An input layer drove E neurons through feedforward weights; reciprocal E→I and I→E weights formed the PING loop, without E→E or I→I recurrence. Loop coupling was 0 or 1.5 the I→E parent mean was 2 times E→I, and both parent standard deviations were 10% of their means. Input parent weights had mean 1.5 and standard deviation 0.3 lower-clamped Gaussian draws were normalised by fan-in, with 95% initially zero and survivors rescaled.
Generate the drive and simulate. Uniform Poisson input drove every channel at the condition’s rate, with seed 42 for input and network initialization. Membranes began at −65 mV and conductances at 0 µS. Exponential-Euler membrane integration used 0.1 ms steps; AMPA and GABA decay constants were 2 and 6 ms, with E/I refractory periods 1.2/0.6 ms. These describe the executed reset holds; the earlier 3/1.5-ms declaration was corrected without rerunning the simulations or changing their measured activity. Full trial recordings included spikes, voltages and conductances.
Measure the drive response. Both loop conditions used 1 trial at each of 2, 5, 10, 20, 40, 70, 100 Hz through 784 input channels for 400 ms, without a discarded transient. Mean per-neuron firing rate was
where identifies E or I, is the population’s total spike count, its neuron count, the full presentation duration in seconds and the rate in hertz. No averaging across independent seeds was performed.
Estimate the population spectrum. The mean E spike trace was demeaned and passed to Welch’s density estimator with one full-trial window.[1] This is a single-window estimate, not an average over independent segments. The largest peak between 5 and 150 Hz was refined by three-bin parabolic interpolation, clamped to half a bin; it was reported only when I spikes were present. This reporting rule is not a test of significant rhythmicity.
Reconstruct signed currents. Each displayed neuron was the first neuron attaining the highest total spike count in its population during that raster trial. A silent E population used neuron zero; a silent I population had no selected-neuron panels. Raster trials used 1024 input channels at 5 Hz for COBA and 45 Hz for PING, unlike the matched-drive sweep.
For the selected neurons, recorded conductances and voltages gave