013 — Training equivalence across Δt regimes

Abstract

First of three Δt-stability notebooks. Asks: “Do the five model variants (standard-snn, snntorch-library, cuba, coba, ping) train to comparable end-state when integrated at $\Delta t = 0.1$ ms vs $\Delta t = 1.0$ ms, starting from matched initial weights?”

The trained checkpoints from this entry are reused by nb014 (eval- $\Delta t$ transport).

Methods

For each of the two training regimes $\Delta t \in \{0.1, 1.0\}$ ms, train every model variant at the same seed and the same MNIST recipe (Poisson-encoded inputs, 200 ms trials, mem-mean readout). Verify that within the CUBANet family (standard-snn, snntorch-library, cuba) the pre-scale weights match bit-exactly at the matched seed; coba and ping use a different class (COBANet) and are independent.

Per-model training is run through the oscilloscope CLI with hyperparameters hardcoded in the runner per the “notebook is the recipe” rule. The single CLI override exposed is --tier.

Results

Per-epoch training curves

Per-epoch train loss and test accuracy for each model variant, with two regimes (dt=0.1 ms and dt=1.0 ms) shown side-by-side. — Train loss and test accuracy per epoch, one curve per model and Δt regime. The qualitative finding: standard-snn and snntorch-library track each other tightly; cuba converges to comparable accuracy with the $(1-\beta)/\Delta t$ drive scaling; coba/ping converge through their own loop dynamics.

Discussion

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

Next steps

nb014 — eval-Δt transport across encoder modes.