This collection asks whether a pyramidal–interneuron network gamma (PING) loop can act as a structural sparsity constraint in a task-trained spiking network. The experiments separate three questions that are easy to muddle: what generates the rhythm, what sets the firing-rate floor, and whether the resulting sparse activity remains useful for classification. They also distinguish a genuine experimental dependency—one entry consuming another entry’s checkpoints or measurements—from a looser conceptual dependency.
Link to exp022. The collection’s operational hub owns the canonical training registry and produces the shared COBA and PING checkpoint banks.
Child experiments
exp024 — Accuracy converges, firing rate does not
Link to exp024. Audits exp022′s per-epoch baseline histories, comparing convergence of classification accuracy and firing rate. Manuscript: supports the rate-attractor interpretation of Figure 3; it is not the plotted image source.
exp025 — PING locks E rate ≈10× below COBA
Link to exp025. Uses exp022′s trained cells for the central comparison between loop-free COBA and gamma-gated PING. Manuscript: Figure 3.
exp037 — PING tolerates 80% dropped spikes but collapses on added noise
Link to exp037. Applies spike-deletion and spike-addition perturbations to exp022′s trained cells. Manuscript: Figure 8.
exp038 — Switching the loop on at inference cuts E rate ≈15×
Link to exp038. Transfers a trained loop-free baseline into a fresh inhibitory architecture, separating the loop’s inference-time effect from training history. Manuscript: Figure 4.
exp041 — E rate is affine in gamma frequency
Link to exp041. Uses exp022′s family to measure how inhibitory decay changes gamma frequency and excitatory rate. Manuscript: supplies the spiking comparison in Figure 2 and the complete Figure 6.
exp044 — Rate floor stable across a 20× Δt sweep
Link to exp044. Reads exp022′s timestep-sweep cells to test whether the rate floor is physical rather than a step-count artefact. Manuscript: Figure 10.
exp048 — Temporal and spatial evidence limits of trained PING
Link to exp048. Loads the canonical trained PING checkpoint and varies presentation time, input rate, and evidence distribution. Manuscript: current source for Figures 11 and 12, pending replacement by exp082.
exp049 — Gradient descent does not preserve a trainable PING loop
Link to exp049. Uses exp022′s initialisation-family cells to test whether training retains or dismantles the recurrent loop. Manuscript: Figure 5.
exp082 — Variable-rate streaming with a spike-rate readout
Link to exp082. Consumes exp022′s variable-rate checkpoint bank. Exp080 and exp081 support its design conceptually but are not execution dependencies. Manuscript: planned replacement source for Figures 11 and 12 after results are complete.
Link to exp023. Establishes the free-running excitatory–inhibitory gamma mechanism without training. Manuscript: Figure 1.
Link to exp047. Separates nominal total coupling from realised per-synapse strength while varying inhibitory-pool size.
Link to exp054. Calibrates a rhythmicity statistic and identifies its low-rate and shared-input failure modes using independent simulations. Manuscript: coupling-plane and spiking panels in the Figure 2 composite.
Link to exp080. Calibrates transformations of sparse pixel intensities into variable Poisson input rates. It conceptually informs exp082 but supplies no runtime artifact.
Link to exp081. Analyses how sparse conductance-driven pixel streams are filtered before and within the PING network. It conceptually informs exp082 but supplies no runtime artifact.