Stochastic Perturbations of Tabular Features for Non-Deterministic Inference with Automunge
share
Injecting gaussian noise into training features is well known to have regularization properties. This paper considers noise injections to numeric or categoric tabular features as passed to inference, which translates inference to a non-deterministic outcome and may have relevance to fairness considerations, adversarial example protection, or other use cases benefiting from non-determinism. We offer the Automunge library for tabular preprocessing as a resource for the practice, which includes options to integrate random sampling or entropy seeding with the support of quantum circuits for an improved randomness profile in comparison to pseudo random number generators. Benchmarking shows that neural networks may demonstrate an improved performance when a known noise profile is mitigated with corresponding injections to both training and inference, and that gradient boosting appears to be robust to a mild noise profile in inference, suggesting that stochastic perturbations could be integrated into existing data pipelines for prior trained gradient boosting models.
READ FULL TEXT
