Gaussian kernel use in ML questioned over predictive uncertainty and conditioning issues

A new paper argues against the widespread use of the Gaussian kernel in machine learning tasks like regression and classification. The authors contend that this kernel, also known as the squared exponential or radial basis function kernel, leads to unrealistically small conditional variances, resulting in catastrophic overconfidence in predictive uncertainty. Furthermore, this kernel's inherent smoothness causes numerical ill-conditioning, necessitating workarounds like nugget terms for practical application. The paper suggests that analytic kernels, in general, should be avoided. AI

IMPACT Challenges the default use of a common kernel, potentially impacting model development and uncertainty quantification in ML applications.

RANK_REASON Academic paper detailing theoretical findings and arguments about a specific machine learning technique. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• classification
• nugget terms
• predictive uncertainty
• radial basis function kernel

• paper
• other

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