014 — Eval-Δt transport across encoder modes

Abstract

Second of three Δt-stability notebooks. Asks: “Can a network trained at one $\Delta t$ run accurately at a different $\Delta t$ at inference? And which spike-resampling mode preserves the input statistics?”

Reads the checkpoints produced by nb013; no retraining.

Methods

For each of nb013’s $5 \text{ models} \times 2 \text{ train regimes} = 10$ trained networks, sweep the evaluation $\Delta t$ over a per-regime grid (fine train- $\Delta t$ → broad grid, coarse train- $\Delta t$ → sparser grid of integer divisors). At each eval- $\Delta t$ , encode the test inputs three different ways:

upsample. Zero-pad the reference spike stream to a finer eval- $\Delta t$ . Eval- $\Delta t \leq$ train- $\Delta t$ only.
downsample. Sum-pool the reference stream to a coarser eval- $\Delta t$ . Eval- $\Delta t \geq$ train- $\Delta t$ only.
resample. Generate a fresh Poisson stream at the target eval- $\Delta t$ . Works in both directions.

Measure test accuracy and mean hidden firing rate at each (model, train- $\Delta t$ , encoder mode, eval- $\Delta t$ ).

Results

Accuracy vs evaluation Δt

Test accuracy vs eval-Δt, panels by training regime, curves by model and encoder mode. — Test accuracy vs eval- $\Delta t$ , one panel per training regime, curves coloured by model and styled by encoder mode. The headline: cuba (with $(1-\beta)/\Delta t$ drive scaling) holds accuracy flat across eval- $\Delta t$ regardless of train- $\Delta t$ ; the snnTorch family sags as eval- $\Delta t$ departs from train- $\Delta t$ , and the sag-point follows train- $\Delta t$ .

Hidden firing rate vs evaluation Δt

Mean hidden firing rate (Hz) vs eval-Δt, panels by training regime, curves by model and encoder mode. — Mean hidden firing rate vs eval- $\Delta t$ . The count-preserving modes (upsample / downsample) keep the input spike count constant across eval- $\Delta t$ but the cuba scaling compensates whereas the snnTorch models do not.

Discussion

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