Post
294
New NeurIPS paper: “Training Data Attribution via Approximate Unrolling”
Ever wondered how individual data points influence AI decisions? 🤔 We explore how specific training data pieces affect machine learning models' behavior, which can be crucial for making AI systems more transparent, trustworthy, and fair.
Our method, SOURCE, bridges the gap between implicit differentiation and unrolling approaches, combining computational efficiency with flexibility making it suitable for non-converged models and multi-stage training pipelines.
📄 Full paper: https://openreview.net/pdf?id=3NaqGg92KZ
Juhan Bae led along with Wu Lin and Roger Grosse.
Supported by the University of Toronto, Vector Institute, NVIDIA, and Anthropic
Ever wondered how individual data points influence AI decisions? 🤔 We explore how specific training data pieces affect machine learning models' behavior, which can be crucial for making AI systems more transparent, trustworthy, and fair.
Our method, SOURCE, bridges the gap between implicit differentiation and unrolling approaches, combining computational efficiency with flexibility making it suitable for non-converged models and multi-stage training pipelines.
📄 Full paper: https://openreview.net/pdf?id=3NaqGg92KZ
Juhan Bae led along with Wu Lin and Roger Grosse.
Supported by the University of Toronto, Vector Institute, NVIDIA, and Anthropic