Utilizing Synthetic Data for VV&C of Machine Learning Applications

Machine Learning (ML) offers a revolutionary method to build algorithms from data. This data-driven methodology has proven to be very effective in progressing fields ranging from natural language processing to autonomy. Simultaneously, the technique has also opened significant challenges in system validation with important issues when used in safety-critical systems. If data forms function, how does one know the generated function is safe? Further, fundamental to the ML approach is the availability of vast amounts of data for training a ML model. For a variety of important applications, the availability of data is challenging.In this paper, we propose a unique computational paradigm which combines the inherent advantages of ML with algorithmic transformational methods. Using this technique, one can generate a large amount of synthetic data which preserve the "signature" of the original dataset while building variations that aid the ML training process. Further, the nature of the algorithmic transformations offers insights into the validation and verification (V&V) scenario generation process. An example is provided of the methodology exercised in a maintenance application for the infrastructure delivering the safety critical Automatic Dependent Surveillance-Broadcast (ADS-B) services in the United States (US).