Oracle ‐ Guided Synthesis of Machine Learning Models Sanjit A. Seshia Professor EECS, UC Berkeley Publication: “Towards Verified Artificial Intelligence,” S. A. Seshia, D. Sadigh, and S. S. Sastry, June 2016. Dagstuhl Seminar March 20, 2018
Correct ‐ by ‐ Construction Design of ML Systems What’s the Spec.??? High ‐ Level Specification Focus in this talk: Synthesizer Use of ML for Perception What to Synthesize? Synthesized ML System S. A. Seshia 2
What’s the (Formal) Specification? 1. System ‐ Level Specification – Captures Application/Context – Need not involve I/O of ML model 2. Component ‐ Level Specification – Robustness to perturbations – Invariance to certain + transformations … – Anything else??? S. A. Seshia 3
Example: Automatic Emergency Braking System (AEBS) Environment Sensor Input Controller Plant Learning ‐ Based Perception System ‐ Level Spec.: (signal temporal logic) G ( dist (ego vehicle, env object) > ) (for all objects) • Goal: Falsification (find counterexamples) • Simulation models of Controller, Plant, Env (e.g. Matlab/Simulink) • Multiple Deep Neural Networks: Inception ‐ v3, AlexNet, SqueezeDet, Yolo, … S. A. Seshia 4
Our Approach: Combine Temporal Logic CPS Falsifier with ML Analyzer Signal Space too large for Spec Error? CPS Falsifier! CPS Model Falsifier (Always) Wrong ML Sensor inputs Region of (images) Uncertainty Perfect ML ML Analyzer Compositional Verification without Compositional Specification! 1. Dreossi, Donze, Seshia, “Compositional Falsification of Cyber ‐ Physical Systems with Machine Learning Components”, NFM 2017. 2. Seshia, “Compositional Verification without Compositional Specification for Learning ‐ Based Systems”, UCB EECS Tech. Report, 2017. S. A. Seshia 5
Machine Learning Analyzer Systematically Explore Region of Interest in the Image (Sensor) Space brightness car z-pos brightness car z-pos Abstraction map car x-pos Systematic Abstract space A Semantic modification space � � Sampling (low ‐ discrepancy sampling) � ✓ ���� ✓ ✓ � ✕ ✕ ✓ Neural network ✓ ✕ � ∈ ����, ����� x Abstract space A ✓ ✕ 6 S. A. Seshia
Sample Result Misclassifications Not of concern Corner case Image But this one is a real hazard! 7 Not trained enough with cars in the middle?
What to Synthesize (of the ML model)? • Training/Test Data • Model Parameters • Hyperparameters • Model Structure • … S. A. Seshia 8
Oracle ‐ Guided Inductive Synthesis (OGIS) Inductive Synthesis : Learning from Examples (ML) Formal Inductive Synthesis : Learn from Examples while satisfying a Formal Specification Key Idea: Oracle ‐ Guided Learning Combine Learner with Oracle (e.g., Verifier) that answers Learner’s Queries query response ORACLE LEARNER [Jha & Seshia, “A Theory of Formal Synthesis via Inductive Learning”, 2015, Acta Informatica 2017.] S. A. Seshia 9
Counterexample ‐ Guided Training of DNNs • CEGIS: Instance of Oracle ‐ Guided Inductive Synthesis • Oracle is CPSML Falsifier used to perform counterexample ‐ guided training of DNNs • Substantially increase accuracy with only few additional examples Learned Classifier DEEP NEURAL FALSIFIER NETWORK (CPS + ML) S. A. Seshia 10
Counterexample ‐ Guided Retraining Blind spot squeezeDet Neural Network (trained on synthetic images using TensorFlow) Example of counterexamples • Precision & Recall improved by more than 10% over standard data augmentation methods [Dreossi, Fremont, Ghosh, Xue, Keutzer, Sangiovanni ‐ Vincentelli, Seshia, 2017, 2018.] S. A. Seshia 11
Summary • Generate “semantic adversarial examples” that violate system ‐ level specification • Compositional Approach without Compositional Specification • Augmenting training set with resulting data (e.g. images) until no more counterexamples • Ongoing/Future Work: – Improving ML analyzer (image synthesizer, space exploration, etc.) – Expanding sensor data (e.g., video, LIDAR) – Learning in/for Human “Towards Verified Artificial Intelligence,” Cyber ‐ Physical Systems S. A. Seshia, D. Sadigh, and S. S. Sastry, 2016. (VeHICaL project) 12
Recommend
More recommend
