New Sobol' index estimators enable sensitivity analysis for large neural networks

Researchers have developed scalable extensions to given-data Sobol' index estimators, enabling variance-based sensitivity analysis for models with a very large number of inputs, such as neural networks with over 10,000 inputs. The new methods include a general definition for arbitrary partitions, a streaming algorithm for batch processing of input-output samples, and a screening heuristic for small indices. These extensions significantly reduce memory requirements compared to existing methods, allowing for the analysis of models with much higher input dimensions while maintaining comparable accuracy and runtime. AI

IMPACT Enables more efficient sensitivity analysis for large-scale AI models, potentially improving their interpretability and robustness.

RANK_REASON Academic paper detailing new methods for sensitivity analysis in machine learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• neural networks
• ScienceCast
• Teresa Portone

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →