Conference Presentation
Kathryn McKinley
- Future applications are expected to increasingly prioritize reasoning about data source trustworthiness over unverified content as the quality of incoming estimates, such as those from agricultural sources, continues to vary significantly from current to future states.
- The system runtime will utilize Bayesian networks to evaluate expressions conditionally rather than immediately, employing a sequential likelihood reweighting algorithm to ensure efficient performance for higher-order inference.
- To facilitate developer adoption, the programming model will be restricted to simplify usage while enabling the implementation of advanced statistical features, with the open-source version anticipated for download soon at Microsoft.
- Users may specify confidence levels to manage false positives and false negatives, and future iterations could allow applications to dynamically weigh evidence quality, such as distinguishing GPS "donut" errors following a Rowley distribution from precise dot locations.
- Data providers will be required to develop specific error models for their estimates to support the system, as traditional assertions will no longer suffice for proving privacy or approximate computing guarantees.
- While researchers remain open to collaboration, the team acknowledges that the rigorous step of proving differential privacy in actual implementations has not yet been completed.
- The system will provide runtime feedback by continuing execution if a probability distribution indicates a condition is likely, or remaining silent if the likelihood is low, with potential syntax modifications dependent on community suggestions.
- Although the team anticipates that many developers will find the intuitive model accessible, they concede that the system may not be suitable for all users.