038 — Switching the loop on at inference cuts E rate ≈10× (trains)

Abstract

Loads a trained COBA network and switches the I-loop on at inference by sweeping ei_strength from 0 to 1. The same feedforward weights that fire at ≈52 Hz without the loop fire at ≈5 Hz with the loop engaged, and accuracy stays within ≈12 pp of the COBA baseline. PING gating is a post-hoc sparsity knob — the architecture, not the training, supplies the gamma dynamics.

Methods

Training recipe (canonical / medium tier):

Parameter	Value
Integration timestep $\Delta t$	0.1 ms
Trial duration $T$	200 ms
MNIST samples (80/20 stratified split of 2000)	1600 train / 400 test ( $\approx$ 2.9% of the 70k-sample MNIST corpus)
Epochs	10

PING and COBA baseline definitions and training recipe are in nb025. This entry runs the COBA → PING I-loop transfer probe at eval time on the trained baselines. The input-rate sweep / f-I curve material that previously lived here has moved to nb023, which is the natural home for “architectural response to drive” content.

Results

COBA → PING I-loop transfer

Inference-only: the trained COBA baseline (seed 42, $\theta_u =$ off) is loaded and ei_strength is overridden at eval time across 11 values from 0 to 1. $W_\text{in}$ and $W_\text{out}$ load from the COBA checkpoint; $W_{ei}$ and $W_{ie}$ are freshly initialised (the COBA checkpoint has these at zero, so skipping the load preserves a functional I-loop). No retraining.

Stacked raster panels showing trained COBA replayed with increasing inference-time ei_strength. — Trained COBA replayed at six inference-time *ei_strength* values; same trial, same weights, fresh I-loop each row.

At ei = 0 the asynchronous-dense COBA pattern persists. By ei ≈ 0.4 the same feedforward weights produce gamma cycles. The PING dynamics come from the inhibitory architecture, not from training — the same weights that fired at 52 Hz without the loop fire at ≈5 Hz with the loop wired up.

Test accuracy vs inference-time ei_strength on COBA-trained weights, with horizontal lines at the trained COBA baseline and at 10% chance. — Test accuracy vs inference-time *ei_strength* on trained COBA weights.

Mean E (hidden) and I (inhibitory) population firing rates per test pass versus ei_strength. — Mean E and I rates across the same *ei_strength* sweep.

Accuracy stays within ≈12 pp of the COBA baseline across the full ei_strength range. E rate falls monotonically from ≈52 Hz (ei = 0) to ≈5 Hz (ei = 1); I rate rises from zero to ≈23 Hz. Suppression is continuous in the loop strength.

Discussion

The I-loop transfers cleanly onto COBA-trained weights: switching on $W^{EI}$ at inference time compresses the E rate by an order of magnitude with a bounded accuracy cost. The PING substrate is therefore not contingent on having been trained with the loop engaged — it can be added post-hoc as a sparsity knob.

Next steps

Repeat the transfer in the opposite direction: train with the I-loop on, then sweep ei_strength down to zero at inference. Symmetric collapse would confirm the loop is a sparsity knob rather than a learned-circuit property.
Quantify the accuracy/rate Pareto frontier traced by the ei_strength sweep and compare to the per-architecture frontier from nb025.