Evolving Abstract Transformers for Gradient-Guided, Adaptable Abstract Interpretation
summary
Proposes a general, parametric transformer that outputs a space of sound outputs (constructed by UPOSE) and then searches it with a differentiable AGG gradient procedure implemented in the AbsEvolve framework. Works for Zones, Octagons and Polyhedra and for linear and quadratic operators, trades precision for time by varying gradient steps, reproduces LP-optimal invariants while reaching them up to 3.2× faster than LP-baseline on NLADigibench, and ships code and artifacts.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 2:30Current transformer limitations
- 7:01Evolving transformers framework
- 10:29Selective dualization
- 13:18Gradient-guided traversal
- 16:06Experimental evaluation
- 18:14Future directions
related talks
11mAbstract Interpretation for Automatic Differentiation
58mBuilding Trust and Safety in Artificial Intelligence with Abstract Interpretation
26mADEV: Sound Automatic Differentiation of Expected Values of Probabilistic Progra...
10mReverse mode ADEV via YOLO: tangent estimators transpose to gradient estimators
23mType-Preserving, Dependence-Aware Guide Generation for Sound, Effective Amortize...
45mScale Learning and Reasoning Across Heterogeneous Gradients and Semantics
25mYou Only Linearize Once: Tangents Transpose to Gradients
18mA Language and Smoothed Semantics for Convergent Stochastic Gradient Descent