msif-isac

Reproduction code for the paper:

Dikmen, İ. C. (2026). Does Nature Know Fourier? Lessons from Echolocating Species for FFT-Free Neuromorphic Radar and ISAC. (Manuscript under review.)

This repository reproduces the 72-cell grand cross-validation (6 MS-IF × 4 H-LIF × 3 datasets), the FFT baseline comparisons, the leave-one-file-out and leave-one-sequence-out audits, and all figures in the paper.

The MS-IF and H-LIF neuron models used here are a subset of the broader dikmen-spiking-neurons library (39 models across the H-LIF, MS-IF, and related families). This repo carries only the models and thresholds that appear in the paper; the full library carries the rest.

What is in the paper

Two orthogonal design axes for neuromorphic radar neurons:

• MS-IF (Multi-Scale Integrate-and-Fire) — what temporal features the neuron listens to. Five specialized models plus a baseline LIF: Dual-τ, Chirp, Phase, DB, Gabor, Baseline.
• H-LIF (Hazard-Based LIF) — how the neuron decides to fire. Four threshold variants: Det, Stoch, BH, TH.

Three real-world datasets (60 GB total):

• FMCW radar interference mitigation (Rock et al., IEEE DataPort 10.21227/1fhk-b416)
• DroneRF (Al-Sa'd et al., Mendeley Data 10.17632/f4c2b4n755.1)
• Xiangyu raw ADC 77 GHz FMCW (Gao et al., IEEE DataPort 10.21227/xm40-jx59)

Installation

Python 3.10+ required. Clone the repo and install:

git clone https://github.com/DrCanD/msif-isac.git
cd msif-isac
pip install -e .

To run tests:

pip install -e ".[dev]"
pytest tests/

Data preparation

The three datasets are not redistributed here. See docs/datasets.md for download instructions and expected directory layout. Default paths are ./data/fmcw/, ./data/dronerf/, ./data/xiangyu/. Override via the config files in configs/.

Reproducing the paper

The four scripts below reproduce the results tables and figures. Stochastic models (Stoch, BH, TH) are 3-run averaged by default. Bootstrap resampling uses 5000 iterations for FMCW/DroneRF and 1000 for Xiangyu.

# 72-cell grand cross-validation (Table 5, Figure 3)
python scripts/run_grand_cv.py --config configs/fmcw.yaml --out results/fmcw.json
python scripts/run_grand_cv.py --config configs/dronerf.yaml --out results/dronerf.json
python scripts/run_grand_cv.py --config configs/xiangyu.yaml --out results/xiangyu.json

# FFT baselines (Table 6, Figure 5)
python scripts/run_fft_baselines.py --all --out results/fft_baselines.json

# LOFO / LOSO audits (Figure 4)
python scripts/run_lofo_audit.py --dataset dronerf --out results/dronerf_lofo.json
python scripts/run_lofo_audit.py --dataset xiangyu --loso --out results/xiangyu_loso.json

# Hyperparameter sweep validation (Section 5.1 footnote)
python scripts/run_hyperparameter_sweep.py --dataset fmcw --cell dual_tau_det

# Figures
python scripts/make_figures.py --all

Total wall-clock on a single modern CPU: roughly 4–6 hours end to end. No GPU required.

Default parameters

H-LIF parameters are fixed across datasets (Section 5.1):

Variant	Parameters
Det	(none)
Stoch	σ = 0.08
BH	λmax = 3, κ = 5, η = 3, δ = 0.1
TH	λmax = 0.5, κ = 1, η = 3, δ = 0.02

Membrane parameters are per dataset, set by a 27-configuration grid search on a calibration subset:

Dataset	βm	θ	refr
FMCW	0.9	0.15	3
DroneRF	0.9	0.05	3
Xiangyu	0.85	0.15	2

MS-IF feature parameters (ftune, τ, barrier thresholds) are set by biological analogy or signal physics and are not optimized.

Citation

If this code helps your work, please cite the paper:

@article{dikmen2026nature,
  author  = {Dikmen, İsmail Can},
  title   = {Does Nature Know Fourier? Lessons from Echolocating Species for FFT-Free Neuromorphic Radar and ISAC},
  journal = {(Manuscript under review)},
  year    = {2026}
}

@software{dikmen2026msifisac,
  author    = {Dikmen, İsmail Can},
  title     = {msif-isac: Reproduction code for 'Does Nature Know Fourier?'},
  version   = {v1.0.0},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19713940},
  url       = {https://doi.org/10.5281/zenodo.19713940},
  license   = {Apache-2.0}
}

and, if you use neurons from the broader library:

@software{dikmen2026snlib,
  author  = {Dikmen, İsmail Can},
  title   = {dikmen-spiking-neurons: A library of 39 spiking neuron models for neuromorphic signal processing},
  year    = {2026},
  url     = {https://github.com/DrCanD/dikmen-spiking-neurons},
  license = {Apache-2.0}
}

License

Apache License 2.0. See LICENSE.

The Apache 2.0 license includes an explicit patent grant to users of this software. A patent application covering the TH-LIF neuron model and the MS-IF neuron family is in preparation with the Turkish Patent and Trademark Office (TÜRKPATENT). The license permits open-source use and redistribution under the terms of Apache 2.0.

Author

İsmail Can Dikmen — Assistant Professor, Electrical and Electronics Engineering, İstinye University, Istanbul, Turkey.

• ORCID: 0000-0002-7747-7777
• Web: Google Scholar
• IEEE Senior Member