Traditionally, data storage systems provide error-correction and data integrity techniques in an independent layer, with the goal of protecting all the data equally regardless of the application. In the context of machine learning systems, this strategy is not appropriate: errors in a few features may prove to be critically important (with respect to the algorithm output), while many errors may have little or no effect. This work takes a different direction: we allow ML algorithms to talk to error-correction schemes, with the goal of making algorithms robust to storage noise. We introduce several novel problems, provide an efficient solution to estimate the noise-induced change in the algorithm output for linear models, and show how to optimize error-correction codes to minimize error effects for fixed overhead.

@inproceedings{SalaRMLW17,
  author    = {Sala, Frederic and Kabir, Shahroze and Dolecek, Lara and Van den Broeck, Guy},
  title     = {Don’t Fear the Bit Flips: Robust Linear Prediction Through Informed Channel Coding},
  booktitle = {ICML 2017 Workshop on Reliable Machine Learning in the Wild},
  month     = Aug,
  year      = {2017},
  url       = "https://starai.cs.ucla.edu/papers/SalaRMLW17.pdf",
  keywords  = {workshop}
}