Kernelized Parameter Sensitivity Atlases for Amortized Conditional Interpretability

Key points

• Gradient-based attribution provides input-specific information about model parameters, but computing it via backpropagation for many samples is expensive, while static importance scores discard input dependence.1
• We derive bounds that separate finite-source from irreducible representation error and give margin-based top-$k$ recovery.1
• Firstly, KPSA identifies parameter groups with high susceptibility to perturbations across vision, forecasting, and protein-modeling setups.1
• The empirics show that compact kernel atlases can store useful input-dependent parameter sensitivity while amortizing repeated gradient attribution, a step toward scalable conditional interpretability.1

Sources (1)

• [1]Kernelized Parameter Sensitivity Atlases for Amortized Conditional Interpretability
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 09:35 PM

  We derive KPSA (Kernelized Parameter Sensitivity Atlas) from a block decomposition of the empirical neural tangent kernel.

  Gradient-based attribution provides input-specific information about model parameters, but computing it via backpropagation for many samples is expensive, while static importance scores discard input dependence.

Extractive summary: sentences quoted from the sources.