Researchers have developed a new method called the Accumulative SGD-Influence Estimator (ACC-SGD-IE) to improve the accuracy of training-data influence estimates. This technique addresses a bias in existing methods by accounting for how earlier data removals affect the impact of subsequent removals. Experiments on various datasets, including tabular, text, and vision data, show that ACC-SGD-IE reduces prediction errors and improves data cleansing effectiveness compared to the standard SGD-Influence Estimator. AI
IMPACT Enhances data cleansing and selection techniques, potentially leading to more robust and efficient model training.
RANK_REASON Academic paper detailing a new methodology for influence estimation in machine learning. [lever_c_demoted from research: ic=1 ai=1.0]
- ACC-SGD-IE
- Accumulative SGD-Influence Estimator
- CIFAR-10
- MNIST database
- SGD-based Leave-One-Out Influence
- SGD-IE
- SGD-Influence Estimator
- Yunxiao Shi
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →