Neural Networks Can Make Worse Decisions When Pooling Data, Study Finds

A new paper explores how neural networks can make worse decisions when combining data from different sources, a phenomenon known as preference reversal. The research identifies that refitting models on pooled data can alter the underlying geometry used to weigh evidence, potentially reversing previously consistent preferences. The study proposes methods to measure and mitigate this issue, including a Gram mismatch metric for pool selection and geometry-oriented regularization during training, along with a three-stage audit to trace decision changes and utility loss. AI

IMPACT This research highlights potential reliability issues in AI systems that aggregate data, suggesting a need for more robust auditing and training methods to ensure consistent decision-making.

RANK_REASON The item is an academic paper detailing a novel finding about neural network behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• alphaXiv
• Case-Based Decision Theory
• CatalyzeX Code Finder for Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Gram Incompatibility
• Gram mismatch
• Hugging Face
• Influence Flower
• ordinary least squares
• Neural Networks
• ScienceCast

• paper
• safety

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