Abstract: Analysts and scientists are increasingly interested in automatically analyzing the semantic contents of unstructured, non-tabular data (videos, images, text, and audio). In order to extract the semantic contents, analysts have turned to machine learning (ML) methods, which can be used in unstructured data analytics systems. Unfortunately, using these ML methods requires expertise to deploy and can be incredibly expensive to execute.
To address these issues, I have built AIDB, a database for allowing users to query unstructured data via SQL. In AIDB, a database administrator specifies mappings between virtual columns that are generated via ML models. The application user can then query the tables in AIDB as with any other SQL database. I have also developed new optimizations to accelerate these ML-based queries via approximations and new query optimization techniques, which can provide up to 300x speedups at 95% accuracy.
Abstract: ML has demonstrated great performance in control and perception, but not safety. But flight control requires analyzable behavior and verifiable safety for certification. In addition to providing a certifiable backup controller, ML-Simplex Architecture creates an analyzable and certifiable safety envelope for ML perception and control. ML operates the UAV with optimized performance whenever it stays with the safety envelope. Or the backup controller takes over. ML Simplex architecture enables ML to safely learn when operating the UAV. Physics expressions are used to augment the NN’s input vector and cost function, resulting in more effective training.