Understanding Black-box Predictions via Influence Functions
This provides a method for interpreting complex models, which is crucial for improving transparency and trust in AI systems, though it builds on existing statistical techniques.
The paper tackles the problem of explaining black-box model predictions by using influence functions to trace predictions back to influential training data, and demonstrates their utility for understanding model behavior, debugging, error detection, and adversarial attacks on linear models and CNNs.
How can we explain the predictions of a black-box model? In this paper, we use influence functions -- a classic technique from robust statistics -- to trace a model's prediction through the learning algorithm and back to its training data, thereby identifying training points most responsible for a given prediction. To scale up influence functions to modern machine learning settings, we develop a simple, efficient implementation that requires only oracle access to gradients and Hessian-vector products. We show that even on non-convex and non-differentiable models where the theory breaks down, approximations to influence functions can still provide valuable information. On linear models and convolutional neural networks, we demonstrate that influence functions are useful for multiple purposes: understanding model behavior, debugging models, detecting dataset errors, and even creating visually-indistinguishable training-set attacks.