Interpreting Deep-Learned Error-Correcting Codes

Deep learning has been used recently to learn error-correcting encoders and decoders which may improve upon previously known codes in certain regimes. The encoders and decoders are learned "black-boxes", and interpreting their behavior is of interest both for further applications and for incorporating this work into coding theory. Understanding these codes provides a compelling case study for Explainable Artificial Intelligence (XAI): since coding theory is a well-developed and quantitative field, the interpretability problems that arise differ from those traditionally considered. We develop post-hoc interpretability techniques to analyze the deep-learned, autoencoder-based encoders of TurboAE-binary codes, using influence heatmaps, mixed integer linear programming (MILP), Fourier analysis, and property testing. We compare the learned, interpretable encoders combined with BCJR decoders to the original black-box code.