Guided Data Generation for Understanding Model Behavior

Abstract

We propose a method for generating distributions over the input space as an inspection tool for understanding trained models. Our framework poses questions of the form ``which inputs would make a trained model exhibit a specified behavior?'' and encodes each question through a guidance function. The generated data provide insights into how the models behave. To showcase our framework, we pose queries such as generating distributions of data where a specified label would be predicted by the model, where two distinct models would disagree, where the output is sensitive to parameter perturbations, and where predictions would be risky. Our method can be applied with a variety of classification and regression tasks and on a range of model types.