Author: Stephen Zorio
Publisher: Amazon: Science
Publication Year: 2020
Summary: In the following article, panelists discuss the tradeoff between accuracy and privacy when dealing with sensitive situations and machine learning (ML) models. They provide a relevant example with the pandemic outbreak, where accurate contact tracing can be more important than respecting user privacy and this highlights two problems: bias and privacy. From their point of view, there is no single correct definition of privacy and the best route to take is differential privacy. Intentionally adding noise to all population-level statistics can mitigate the risks of precisely identifying individuals, since no number is precise but a rough estimate. The solution to address bias is not to eliminate bias from the data but to tune the ML models to standardize the bias (if any). There can be useful patterns that can crop up based on race, for example, but they cannot be immediately assumed as being biased against a certain race. Addressing bias and avoiding features susceptible to bias are not synonymous to each other.
