005 — PING Integration Step Video

Abstract

Gamma has a period of \sim15–30 ms, so a coarse integration step dt can’t resolve the E→I→E round-trip. With the Poisson input overdrive pinned well past the PING threshold from nb003 (so the rhythm would be robust at fine dt), this notebook sweeps dt — isolating the timestep as the only knob that can break the oscillation. Companion sweeps: nb003 (drive) and nb006 (E→I coupling); companion dt-stability training experiment is nb013. See The Notebooks for how this entry’s runner/artifact/figure triple is wired up.

Methods

A scan over the oscilloscope video subcommand with stim-overdrive held at 10× (firmly inside the stable-PING regime mapped by nb003) while dt sweeps 0.05 → 2 ms over 300 frames. All knobs are hardcoded literals in src/pinglab/notebooks/nb005.py per the runner contract.

Parameter Value
Setup
Model ping
Input mnist d0 s0
Seed 42
Architecture
N_E / N_I 512 / 128
T 600 ms
Stimulus & scan
Stim window 200–300 ms
Fixed overdrive 10×
dt scan 0.05–2 ms
Frames / FPS 300 / 30
Provenance
Tier large
Elapsed 24m 39s
Run ID r002
Git SHA ?

Results

Figure 1. Integration-step sweep (scan_dt.mp4)

Each frame is a fresh 600 ms sim on MNIST digit 0, sample 0 with stim-overdrive held at 10×. dt sweeps 0.05 → 2 ms over 300 frames.

Stable until about dt = 1.65 ms; beyond that the whole population saturates after PING onset.

Population rates

WindowE (Hz)I (Hz)
Pre-stim6.80.0
In-stim20.554.2
Post-stim5.61.6

Discussion

TODO: discussion paragraph — write what the results above mean for the project.

Next steps

The breakdown around dt ≈ 1.65 ms sets a hard ceiling on the timestep any downstream training entry can use without losing the rhythm. nb013 tests whether this same ceiling holds end-to-end through training — i.e. whether a model trained at dt near the limit still produces PING at inference. The breakdown mode here (saturation, not collapse) suggests the failure is integrator instability rather than loss of the rhythm itself; a side-by-side with explicit-Euler vs the current scheme would isolate that.

Appendix

Reproduction

uv run src/pinglab/notebooks/nb005.py