Autoencoder Training Data for Non-Differentiable Communication Channels

Ognjen Jovanovic from the Technical University of Denmark presents a dataset for training autoencoders using a gradient-free method based on the cubature Kalman filter. The data is used to perform geometric constellation shaping on differentiable channels and to test robustness on non-differentiable channels including laser phase noise and additive white Gaussian noise. The dataset is associated with an Open Access paper on the paperswithcode platform.

Use Cases

• Training gradient-free autoencoders based on the cubature Kalman filter method described.
• Evaluating geometric constellation shaping performance on differentiable communication channels.
• Testing model robustness to residual phase noise in non-differentiable channels as mentioned in the description.
• Comparing autoencoder performance against standard modulation schemes like Quadrature Amplitude Modulation.

Strengths

• Dataset is associated with a peer-reviewed research paper from the Technical University of Denmark.
• The description provides a clear methodological context for gradient-free training using the cubature Kalman filter.
• Research validates performance against standard schemes like Quadrature Amplitude Modulation and Iterative Polar Modulation.

Limitations

• Row count and dataset size are unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.

Provenance

Source

Technical University of Denmark

Collection Method

Likely generated via numerical simulation for validating a gradient-free training method, as described in the associated paper.