AE-ngComms-code

Reproducible research code for the experimental study in:

The Role of Autoencoders in Next-Generation Communication Systems: Architectures and Domain-Specific Paradigms

David Carrascal, Javier Diaz-Fuentes, Elisa Rojas, Joaquín Álvarez-Horcajo, and José M. Arco.

The repository evaluates how block length, latent dimension, network depth, and encoder-decoder asymmetry affect absolute-temperature reconstruction, nominal compression, and a compute-derived energy proxy. The study compares fully connected symmetric autoencoders (AEs), asymmetric autoencoders (AAEs), Rice-Golomb coding, and Gorilla XOR coding.

Requirements

• Python 3.10 or newer.
• A CPU is sufficient. CUDA is used automatically when available.
• The full 50-seed experiment is computationally expensive; use --quick first to verify an installation.

Create an isolated environment and install the project:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[plot,test]"

Run the test suite and validate the archived tables:

python -m pytest
python scripts/validate_published_results.py

Reproducing the experiments

A smoke test runs one small configuration from each selected sweep:

ae-ngcomms-run \
  --sweeps b c d rice gorilla \
  --quick

The complete article protocol uses 50 seeds, at most 300 epochs, and early stopping after 15 validation epochs:

ae-ngcomms-run \
  --sweeps b c d rice gorilla \
  --seeds 50 \
  --epochs 300 \
  --patience 15

New tables are written to results/reproduced/. The exact tables used in the manuscript remain immutable in results/published/.

The sweep identifiers correspond to:

Sweep	Variables	Fixed settings
b	Depth index 1–5; AE versus AAE	n=100, m=25
c	n={128,256,512,1024}, R={2,4,8,16,32}	Depth index 2
d	m={2,4,8,16,32,64,128}	n=256, depth index 2
rice	Rice-Golomb over first differences	Non-overlapping test windows
gorilla	Gorilla XOR over absolute float32 values	Non-overlapping test windows

Regenerate the result figures from either the published or reproduced tables:

ae-ngcomms-plot \
  --results-dir results/published \
  --output-dir figures/reproduced

Experimental protocol

The included trace has 115,235 Caples Lake N7 air-temperature observations. The deterministic chronological split contains 79,710 training, 9,220 validation, and 26,305 test observations, with 2017 reserved for testing. Training windows use strides 10, 27, and 33; validation and test windows do not overlap.

Each AE window is converted to first differences and independently min-max normalized. The decoder output is denormalized and cumulatively summed using the stored reference value. Hidden widths are geometrically spaced, hidden layers use ELU, the latent projection uses SELU, and the output uses a sigmoid. Training minimizes normalized-domain MSE with Adam at a learning rate of 2e-3; reported MAE is measured after reconstructing absolute temperature.

For the asymmetric topology at depth index h, the encoder receives floor((h-1)/2) hidden layers and the decoder receives the remainder. For the symmetric topology, both branches receive h hidden layers. Linear-layer MACs are counted analytically, and the energy proxy uses 4.6 pJ/MAC.

Nominal AE compression is R=n/m. A transmitted latent vector also needs three floating-point values per window (minimum difference, maximum difference, and reference), so the corresponding value-count ratio is n/(m+3). The archived sweep-B table retains its historical ratio column for compatibility and also records dimensions from which either definition can be recovered.

Repository layout

• src/ae_ngcomms/: maintained models, data pipeline, codecs, training, experiments, and plotting code.
• scripts/: experiment/plot wrappers, result validation, and architecture figure generation.
• data/: the exact input trace and its provenance.
• results/published/: exact CSV tables used in the article.
• figures/published/: final vector result figures.
• tests/: model, data, codec, and result-integrity tests.

Exploratory scripts, superseded experiments, generated model weights, platform-specific dependency lists, and the vendored PlotNeuralNet checkout from the development repository are intentionally excluded.

Data

The included trace is a subset of the American River Hydrologic Observatory dataset:

Bales, R.; Cui, G.; Rice, R.; Meng, X.; Zhang, Z.; Hartsough, P.; Glaser, S.; Conklin, M. (2020). Snow depth, air temperature, humidity, soil moisture and temperature, and solar radiation data from the basin-scale wireless-sensor network in American River Hydrologic Observatory (ARHO) [Dataset]. Dryad. https://doi.org/10.6071/M39Q2V

The dataset is available under CC0. See data/README.md and NOTICE.

Research-code notice

This code supports source-compression experiments over a noiseless latent path. The AE ratio is not a bit-rate measurement, the lossless baselines are not rate-matched to the learned representation, and the energy values are MAC-derived proxies rather than hardware measurements. Validate quantization, serialization, memory, radio, and channel costs before drawing deployment conclusions.

Citation

Please cite the release-specific Zenodo record and the associated article when available. Machine-readable metadata are provided in CITATION.cff.

Carrascal, D., Diaz-Fuentes, J., Rojas, E., Álvarez-Horcajo, J.& M. Arco, J. (2026). The Role of Autoencoders in Next-Generation Communication Systems: Architectures and Domain-Specific Paradigms (Version 1.0.2) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.21721615

License

The software is licensed under the Apache License 2.0. See LICENSE. The data subset is distributed under its original CC0 terms and is not relicensed under Apache-2.0.