Reports
Output Evidence Against Time and Gamma Cycles
Paired inhibitory-decay probes of output evidence, matched-coordinate accuracy and accuracy–deadline–spike-cost frontiers across input encoding rates.
Modulating Inhibitory Decay, Leak and Capacitance in PING networks
Paired-input sweeps of inhibitory decay, membrane leak and capacitance, measuring spectral frequency, excitatory participation, spike rate and rhythmic contrast.
Mean-Field Analysis of PING Bifurcations
An independent four-variable mean-field analysis identifies a Hopf bifurcation, sampled dynamics consistent with a supercritical transition, and decreasing onset frequency with slower inhibition.
COBA and PING Dynamics, Discretisation, and Learning
A textbook derivation of the conductance-based COBA and PING equations, their discrete simulation, and their training by surrogate-gradient BPTT, voltage-gradient damping, gradient clipping, AdamW, and Dale projection.
Manuscript
A manuscript scaffold connecting PING circuit dynamics, low-rate task performance, cycle participation, perturbation sensitivity and continuous-stream classification.
Susin and Destexhe (2021)
Independent excitatory afferent drive strengthens population bursting in a calibrated conductance-based E/I network.
Spike-Count Classification in a Continuous Stream
A multi-seed study of spike-count classification across input rates and presentation durations.
How Pixel Features Respond to Input Rate
Direct simulations and a simpler theory test how sparse input events become average voltage features.
Decoder Accuracy Improves with Input Rate
Direct-simulation decoder calibration of the input-rate range for later variable-rate PING training.
Pinglab Rythmicity Metric
Lobe–trough contrast across untrained PING coupling strengths, with private- and shared-input null controls.
Training Recurrent Weights Weakens PING Rhythmicity
Trainable recurrent conductances produced lower reference-image rhythmicity and higher excitatory firing than the frozen PING control; outcomes depended on initialization.
Pool-Size Effects Depend on Synaptic Scaling
Paired controls separate fixed summed I→E coupling from fixed expected synaptic strength as the inhibitory pool grows.
One Spike per Gamma Cycle
Pooled and equally weighted network distributions support predominantly one-spike gamma-cycle participation across 18 trained networks, with exceptions.
Firing Rate Across the Timestep Sweep
Compares final-epoch firing rate, classification accuracy and illustrative rasters across a twelvefold integration-timestep sweep.
Inhibitory Replay Perturbations Change Excitatory Firing
Recorded inhibitory spike streams produced contrasting excitatory responses under independent-spike and fixed-window replay perturbations, but the experiment does not isolate synchrony.
Firing Rate Tracks Gamma Frequency
Across PING networks trained at different inhibitory decay times, compare excitatory firing rate with gamma frequency and test accuracy.
Switching On the Inhibitory Loop
Enabling an inhibitory loop after feedforward training reduced excitatory firing but lowered classification accuracy; the experiment does not isolate a benefit of gamma timing.
Dropped Spikes vs Added Noise
Both trained networks tolerated substantial spike deletion, but PING accuracy fell more sharply under added spikes. The perturbations changed both recurrent feedback and readout input, so they do not isolate gamma gating.
Accuracy and Firing Rate With and Without Inhibition
Reused MNIST networks compare accuracy and excitatory firing rates with and without an inhibitory loop. Different gradient damping limits causal attribution to gamma timing.
Accuracy Plateaus While Firing Rate Rises
Audits validation accuracy and firing-rate stability in unregularised PING and COBA training histories.
Turning the PING Loop On
PING stripped to its biophysical fundamentals: free-running activity, population spectra and rate responses with the excitatory–inhibitory loop off and on.
Training Runs
Seven controlled training families, their retained checkpoint bank, validation learning curves, and raster diagnostics.