Rate floor stable across a 20× Δt sweep

Rate floor stable across a 20× Δt sweep#

exp044 · 2 June 2026 · pdf

The trained networks this entry uses are produced once in the shared training hub, exp022 (Training), and reused here rather than retrained.

Abstract#

The Δt audit asks whether the exp025 headline E rate is a physical (Hz) property of the trained network or an artefact of the integration timestep. The rate stays within a 9–14 Hz band across a 20× Δt sweep, accuracy holds at 90.4–91.4%, and the gamma cycle period in physical ms is invariant. The rate is not fully Δt-independent, though: it rises monotonically with the timestep, from ≈ 9.2 Hz at Δt = 0.05 ms to ≈ 13.4 Hz at Δt = 1 ms, so a coarser step inflates the rate while leaving accuracy and cycle period intact.

Method#

Parameter	Value
Integration timestep $Δ 𝑡$	0.05–1.0 ms (swept)
Trial duration $𝑇$	200 ms
MNIST samples (80/20 stratified split of 7000)	5600 train / 1400 test (≈ 10% of the 70k-sample MNIST corpus)
Epochs	50

The exp022 hub trains one PING per $Δ 𝑡 \in {0.05, 0.1, 0.25, 0.5, 1.0}$ ms × seed $\in {42, 43, 44}$ = 15 cells; this entry loads them and evaluates. Total physical time $𝑇 = 200$ ms is held constant (step count varies 4000 → 200). Batch size 64 throughout, smaller than exp025′s 256 but matched across the sweep so per-step compute and memory stay comparable, and the Δt = 0.05 cells (4000 timesteps × $𝑁_{𝐸}$ × $𝑁_{𝐼}$ ) fit in a single A100. All other PING recipe parameters held to exp025.

Run inference on the test set; report mean E rate (Hz), accuracy, and a single-trial raster from seed 42 per Δt for visual cycle-period inspection.

Results#

Hidden E rate and test accuracy against integration timestep; accuracy is flat while the E rate rises gently with coarser Δt. — Figure 1: Hidden E rate (black) and test accuracy (red) as Δt varies 20×; markers are per-Δt means over three seeds. Accuracy holds near 91% while the E rate rises monotonically from ≈ 9.2 Hz at Δt = 0.05 ms to ≈ 13.4 Hz at Δt = 1 ms.

Single-trial E and I spike rasters at five integration timesteps, plotted against physical time; the burst cadence lines up across all five. — Figure 2: Single-trial rasters at each Δt, x-axis in *physical* ms (not steps). All five panels show E (black) and I (red) bursts locked to the same gamma cadence. The cycle physics is Δt-invariant.

Per-cell test accuracy and E rate versus epoch, coloured by Δt; accuracy converges early while the E rate is still climbing at epoch 50. — Figure 3: Top: test accuracy converges by epoch ≈ 10–15 across all Δt. Bottom: test E rate has *not* converged at epoch 50; every curve is still rising, ordered by Δt. The Δt-dependent rate is a snapshot at epoch 50, not a fixed-point ceiling.