A cross-domain double dissociation of neuron-task matching in pure spiking networks

Experiment code for the paper of the same name (Dikmen & Karadağ, submitted to Neural Networks, Elsevier, 2026). The central claim is that at matched capacity a network whose single neuron type is matched to the task beats a heterogeneous mix of neuron types, and that this advantage flips across domains: the neuron that wins on micro-Doppler radar (DIAT-µSAT) is near the bottom on spoken-digit audio (Spiking Heidelberg Digits), while another neuron shows the opposite pattern. That crossed pattern is the double dissociation.

All networks here are pure spiking: the signal passed between layers is spikes, with no batch or layer normalisation and no softmax gating inside the network. Per-type firing thresholds are calibrated once at initialisation so no layer starts dead or saturated, after which firing runs free and is logged as an efficiency metric.

Layout

src/mstask/            installable package — the shared, verified-equivalent harness
  __init__.py            ensures the spiking-neuron library is importable
  core.py                models, threshold calibration, capacity helpers, metrics, registry
experiments/           one runnable script per experiment (imports from mstask.core)
  01_shd_probe.py              harness probe (SHD T=12), not a paper table
  02_diat_ablation.py          HEADLINE matched-capacity ablation — Section 4.1, Figure 3
  03_best_model_search.py      best-model search — Section 4.2, Table 2
  04_best_model_search_full.py search + compact-CNN reference + op-count proxy — Table 2
  05_shd_bakeoff.py            SHD audio bake-off, speaker-mixed split — Section 4.3, Figure 4
  06_close_2x2.py              DIAT bake-off that closes the dissociation — Section 4.3, Figure 5
  07_shd_disjoint.py           SHD official speaker-disjoint re-run — Section 3.3, Table 3
results/               locked run summaries that back the paper tables (full per-seed scores)
figures/               notes on regenerating the manuscript figures from results/

The experiment scripts were factored out of the original single-cell Colab notebook that produced the paper. The parts that are identical across all experiments (the spiking models, threshold calibration, capacity solver, metrics and the dataset registry) were lifted verbatim into mstask.core after an abstract-syntax-tree comparison confirmed they differ only in formatting, never in computation. The parts that genuinely differ between experiments — the significance test, the training loop, the model builder and the data loaders — were left inside each script so the numbers each one produces stay exactly what they were.

Install

pip install -e .

This installs the mstask package and its dependencies, including the physics-motivated spiking-neuron library (dikmen-spiking-neurons, Zenodo 10.5281/zenodo.20110833), which provides the Doppler, Chirp, Dual-tau, STFT, Phase, Beam and CrossInhib units through NeuronRegistry. Importing mstask also installs that library from GitHub on first use if it is missing, so the scripts also run as-is when pasted into a fresh Colab cell.

Data

The datasets are not redistributed here.

• DIAT-µSAT micro-Doppler (radar): a folder with X.npy [N, 64, 64] and y.npy, six classes, spectrogram-as-sequence with T = 64.
• Spiking Heidelberg Digits (audio): twenty spoken digits from zenkelab.org. Experiment 07 downloads the official speaker-disjoint split and bins it to T = 12; the mixed-split cache used by 01 and 05 is a re-split .pt dictionary.

Point the scripts at your copies with environment variables (the author's Colab paths are kept as defaults for provenance):

export RESEARCH_ROOT=/path/to/research        # run/output root
export DIAT_DATA=/path/to/DIAT_uSAT/processed # DIAT folder with X.npy + y.npy
export SHD_CACHE=/path/to/shd_T12_m200_v2.pt  # mixed-split SHD cache

Run

python experiments/02_diat_ablation.py   # headline ablation
python experiments/07_shd_disjoint.py    # official speaker-disjoint audio ranking

Each script trains three seeds (42, 123, 999), selects on validation-smoothed macro-F1, tests once at the best epoch, and writes a summary.json into its run folder. The headline and disjoint scripts resume from a checkpoint if interrupted. Set REDUCED=1 for a fast reduced configuration where supported.

A note on the environment: the scripts were checked here with py_compile and a static undefined-name pass, but a full end-to-end training run needs a GPU and the datasets, so run experiments/01_shd_probe.py once to confirm the harness on your machine before trusting a long run.

Results

results/ holds the verified run summaries with per-seed scores, so the means, standard deviations and paired-bootstrap statements can be re-derived without re-training. shd_disjoint_summary.json backs Table 3 (Dual-tau-LIF 0.671 top, CrossInhib-LIF 0.635, Doppler-LIF sixth at 0.451; Spearman between the mixed and disjoint rankings 0.76). cnn_frontier_summary.json backs Table 2 (Doppler-LIF 0.924 / 475k, Vanilla 0.857, Parametric 0.852, compact-CNN reference 0.932 / ~24k); its energy fields are an operation-count proxy, not an energy claim.

Citation

If you use this code, please cite the paper, this repository (DOI 10.5281/zenodo.20763581) and the neuron library; see CITATION.cff. Licensed under MIT.