SHF: Small: Distribution-aware Testing for Neural Networks

University Of Virginia Main Campus, Charlottesville VA

Investigators

Abstract

Machine learning aims to learn the definition of a software implementation from a set of examples. For instance, from a large set of images from the forward facing camera in an automobile such approaches can learn to detect when a pedestrian is near the road. This method is attractive because traditional software development approaches have a hard time with this kind of problem specification -- consider the challenge of precisely describing what a pedestrian looks like given all the variations in their clothing, what they are carrying, their location and pose, whether they are partially obscured, partially lit, etc. Using machine learning, new capabilities can thus be used to deploy potentially better systems. The challenge is how to determine that these new capabilities work as expected. For traditional software, decades of research has yielded sophisticated frameworks for testing software. This project will adapt the strategies from that body of research to test machine learning models. In this way the project aims to improve the safety and quality of advanced machine-learning-based software systems, which will have positive impacts on people and organizations who depend on those systems. The project will exploit learned models of a target data distribution to drive testing of neural networks that are trained on samples from that distribution. The methods that are developed will exploit generative models that exhibit high-precision and recall of the target distribution and that have a well-defined mathematical structure to their latent space. By targeting this latent space, distribution-aware neural-network testing techniques can leverage its reduced dimensionality, relative to the neural network's input domain, to assess the adequacy of a given test suite and to generate new and valuable tests. In addition, the project will develop a rigorous mutation-based framework for evaluating the fault-detection effectiveness of neural network testing methods and to employ that framework in evaluating the benefits of distribution-aware testing methods. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.