Oracle Guided Synthesis of Machine Learning Models Sanjit A. Seshia - - PowerPoint PPT Presentation

Oracle‐Guided Synthesis of Machine Learning Models

Sanjit A. Seshia

Professor EECS, UC Berkeley

Dagstuhl Seminar March 20, 2018

Publication: “Towards Verified Artificial Intelligence,”

• S. A. Seshia, D. Sadigh, and S. S. Sastry, June 2016.

Correct‐by‐Construction Design of ML Systems

• S. A. Seshia

2

High‐Level Specification  Synthesized ML System Synthesizer

What’s the Spec.??? What to Synthesize?

Focus in this talk: Use of ML for Perception

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)

• S. A. Seshia

4

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, …

Controller Plant Environment Learning‐Based Perception Sensor Input

Our Approach: Combine Temporal Logic CPS Falsifier with ML Analyzer

5

CPS Falsifier ML Analyzer

Spec Model Region of Uncertainty Sensor inputs (images) Error?

Compositional Verification without Compositional Specification!

• S. A. Seshia

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.

Signal Space too large for CPS Falsifier!

Perfect ML (Always) Wrong ML

Machine Learning Analyzer

6

Systematically Explore Region of Interest in the Image (Sensor) Space

Semantic modification space

• brightness

car z-pos

Abstraction map

brightness car z-pos car x-pos

Abstract space A

• S. A. Seshia

x

Abstract space A

• Neural network

∈ ,

• ✓

✓

✕ ✕ ✕

✓ ✓ ✓ ✓

✕

Systematic Sampling (low‐discrepancy sampling)

Sample Result

7

Misclassifications Corner case Image But this one is a real hazard!

Not trained enough with cars in the middle?

Not of concern

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

• S. A. Seshia

9

[Jha & Seshia, “A Theory of Formal Synthesis via Inductive Learning”, 2015, Acta Informatica 2017.]

Key Idea: Oracle‐Guided Learning

Combine Learner with Oracle (e.g., Verifier) that answers Learner’s Queries

LEARNER ORACLE

query response

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

• S. A. Seshia

10

DEEP NEURAL NETWORK FALSIFIER (CPS + ML)

Learned Classifier

Counterexample‐Guided Retraining

• Precision & Recall improved by more than 10%
• ver standard data augmentation methods
• S. A. Seshia

11

squeezeDet Neural Network (trained on synthetic images using TensorFlow)

Blind spot

Example of counterexamples

[Dreossi, Fremont, Ghosh, Xue, Keutzer, Sangiovanni‐Vincentelli, Seshia, 2017, 2018.]

Summary

• Ongoing/Future Work:

– Improving ML analyzer (image synthesizer, space exploration, etc.) – Expanding sensor data (e.g., video, LIDAR) – Learning in/for Human Cyber‐Physical Systems (VeHICaL project)

12

“Towards Verified Artificial Intelligence,”

• S. A. Seshia, D. Sadigh, and S. S. Sastry, 2016.
• 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