Dependent Coeffects for Local Sensitivity Analysis
summary
Introduces Local Fuzz, a type system using dependent coeffects that computes upper bounds on local sensitivity (sensitivity at point x within radius k), giving finite bounds where Fuzz reports infinity (e.g., square function sensitivity = 2x·k). Provides denotational semantics in extended premetric spaces via a dependently graded comonad, soundness proof that typings yield local-sensitivity bounds, and typing rules for primitives (multiplication, quartiles). Demonstrates a 95% interval algorithm whose quartile-based sensitivity is finite and then privatizes it with less Laplace/PT R noise than global-sensitivity methods.
AI-written summary · about
chapters (10)
- 0:00Dependent coeffects intro
- 1:02Global sensitivity in DP
- 4:54Type systems for sensitivity
- 7:44From global to local
- 9:32Type system design
- 14:50Semantics of coeffects
- 16:33Global sensitivity case studies
- 18:00Quartile interval case study
- 20:54Summary and future work
- 22:20Question and answer session
related talks
16mSensitivity by Parametricity
29mA Pre-Expectation Calculus for Probabilistic Sensitivity
24mSemantics of Probabilistic Programs using S-Finite Kernels in Coq
45mTractability And Overparametrization In Some Simple Models
11mLocal algebraic effect theories
56mSparsifying Generalized Linear Models
15mSafe Couplings: Coupled Refinement Types
32mLocal Statistics, Semidefinite Programming, and Community Detection