Class based Influence Functions for Error Detection
This addresses a reliability issue in anomaly detection for large-scale datasets, though it is incremental as it builds on existing IF methods.
The paper tackled the instability of influence functions (IFs) when applied to deep networks, particularly when data points belong to different classes, and developed a solution using class information that significantly improves performance and stability without extra computational cost.
Influence functions (IFs) are a powerful tool for detecting anomalous examples in large scale datasets. However, they are unstable when applied to deep networks. In this paper, we provide an explanation for the instability of IFs and develop a solution to this problem. We show that IFs are unreliable when the two data points belong to two different classes. Our solution leverages class information to improve the stability of IFs. Extensive experiments show that our modification significantly improves the performance and stability of IFs while incurring no additional computational cost.