دانلود کتاب Deep Neural Networks and Data for Automated Driving: Robustness, Uncertainty Quantification, and Insights Towards Safety

فهرست مطالب :

Foreword Preface About This Book Introduction Contents About the Editors Safe AI—An Overview Inspect, Understand, Overcome: A Survey of Practical Methods for AI Safety 1 Introduction 2 Dataset Optimization 2.1 Outlier/Anomaly Detection 2.2 Active Learning 2.3 Domains 2.4 Augmentation 2.5 Corner Case Detection 3 Robust Training 3.1 Hyperparameter Optimization 3.2 Modification of Loss 3.3 Domain Generalization 4 Adversarial Attacks 4.1 Adversarial Attacks and Defenses 4.2 More Realistic Attacks 5 Interpretability 5.1 Visual Analytics 5.2 Intermediate Representations 5.3 Pixel Attribution 5.4 Interpretable Proxies 6 Uncertainty 6.1 Generative Models 6.2 Monte Carlo Dropout 6.3 Bayesian Neural Networks 6.4 Uncertainty Metrics for DNNs in Frequentist Inference 6.5 Markov Random Fields 6.6 Confidence Calibration 7 Aggregation 7.1 Ensemble Methods 7.2 Temporal Consistency 8 Verification and Validation 8.1 Formal Testing 8.2 Black-Box Methods 9 Architecture 9.1 Building Blocks 9.2 Multi-task Networks 9.3 Neural Architecture Search 10 Model Compression 10.1 Pruning 10.2 Quantization 11 Discussion References Recent Advances in Safe AI for Automated Driving Does Redundancy in AI Perception Systems Help to Test for Super-Human Automated Driving Performance? 1 Introduction 2 How Much Data is Needed for Direct Statistical Evidence of Better-Than-Human Driving? 3 Measurement of Failure Probabilities 3.1 Statistical Evidence for Low Failure Probability 3.2 Test Data for Redundant Systems 3.3 Data Requirements for Statistical Evidence of Low Correlation 4 Correlation Between Errors of Neural Networks in Computer Vision 4.1 Estimation of Reliability for Dependent Subsystems 4.2 Numerical Experiments 5 Conclusion and Outlook References Analysis and Comparison of Datasets by Leveraging Data Distributions in Latent Spaces 1 Introduction 2 Related Works 3 Building Blocks of Our Approach 3.1 Variational Autoencoder DNN 3.2 Methods for Novelty Detection 3.3 Dataset Comparison Approach 4 Experimental Setup 4.1 Datasets 4.2 Experiments 4.3 VAE Hyperparameters 4.4 Training 5 Experimental Results and Discussion 5.1 Results 5.2 Discussion 6 Conclusions References Optimized Data Synthesis for DNN Training and Validation by Sensor Artifact Simulation 1 Introduction 2 Related Works 3 Methods 3.1 Sensor Simulation 3.2 Dataset Divergence Measure 3.3 Datasets 4 Results and Discussion 4.1 Sensor Parameter Extraction 4.2 Sensor Artifact Optimization Experiment 4.3 EMD Cross-evaluation 5 Conclusions References Improved DNN Robustness by Multi-task Training with an Auxiliary Self-Supervised Task 1 Introduction 2 Related Works 3 Multi-task Training Approach 4 Evaluation Setup 5 Experimental Results and Discussion 6 Conclusions References Improving Transferability of Generated Universal Adversarial Perturbations for Image Classification and Segmentation 1 Introduction 2 Related Works 2.1 Environment Perception for Automated Vehicles 2.2 Adversarial Attacks 3 Method 3.1 Mathematical Notations 3.2 Method Introduction and Initial Analysis 3.3 Non-targeted Perturbations 3.4 Targeted Perturbations 4 Experiments on Image Classification 4.1 Experimental Setup 4.2 Non-Targeted Universal Perturbations 4.3 Targeted Universal Perturbations 5 Experiments on Semantic Segmentation 5.1 Experimental Setup 5.2 Non-targeted Universal Perturbations 5.3 Targeted Universal Perturbations 6 Conclusions References Invertible Neural Networks for Understanding Semantics of Invariances of CNN Representations 1 Introduction 2 Background 3 Method 3.1 Recovering the Invariances of Deep Models 3.2 Interpreting Representations and Their Invariances 3.3 Implementation Details 4 Experiments 4.1 Comparison to Existing Methods 4.2 Understanding Models 4.3 Effects of Data Shifts on Models 4.4 Modifying Representations 4.5 Evaluation Details References Confidence Calibration for Object Detection and Segmentation 1 Introduction 2 Related Works 3 Calibration Definition and Evaluation 3.1 Definitions of Calibration 3.2 Measuring Miscalibration 4 Position-Dependent Confidence Calibration 4.1 Histogram Binning 4.2 Scaling Methods 5 Experimental Evaluation and Discussion 5.1 Object Detection 5.2 Instance Segmentation 5.3 Semantic Segmentation References Uncertainty Quantification for Object Detection: Output- and Gradient-Based Approaches 1 Introduction 2 Related Work 3 Methods 3.1 Uncertainty Quantification Protocols 3.2 Deep Object Detection Frameworks 3.3 Output-Based Uncertainty: MetaDetect 3.4 Gradient-Based Uncertainty for Object Detection 4 Experimental Setup 4.1 Databases, Models, and Metrics 4.2 Implementation Details 4.3 Experimental Setup and Results References Detecting and Learning the Unknown in Semantic Segmentation 1 Introduction 2 Anomaly Detection Using Information and Entropy 3 Related Works 3.1 Anomaly Detection in Semantic Segmentation 3.2 Incremental Learning in Semantic Segmentation 4 Anomaly Segmentation 4.1 Methods 4.2 Evaluation and Comparison of Anomaly Segmentation Methods 4.3 Combining Entropy Maximization and Meta Classification 5 Discovering and Learning Novel Classes 5.1 Unsupervised Identification and Segmentation of a Novel Class 5.2 Class-Incremental Learning 5.3 Experiments and Evaluation 5.4 Outlook on Improving Unsupervised Learning of Novel Classes References Evaluating Mixture-of-Experts Architectures for Network Aggregation 1 Introduction 2 Related Works 3 Methods 3.1 MoE Architecture 3.2 Disagreements Within an MoE 4 Experiment Setup 4.1 Datasets and Metrics 4.2 Topologies 4.3 Training 5 Experimental Results and Discussion 5.1 Architectural Choices 5.2 MoE Performance 5.3 Disagreement Analysis References Safety Assurance of Machine Learning for Perception Functions 1 Introduction 2 Deriving Acceptance Criteria for ML Functions 2.1 Definition of Risk According to Current Safety Standards 2.2 Definition of Safety Acceptance Criteria for the ML Function 3 Understanding the Contribution of Safety Evidence 3.1 A Causal Model of Machine Learning Insufficiencies 3.2 Categories of Safety Evidence 4 Evidence Workstreams—Empowering Experts from Safety Engineering and ML to Produce Measures and Evidence 5 Examples for Evidence 5.1 Evidence from Confirmation of Residual Failure Rates 5.2 Evidence from Evaluation of Insufficiencies 5.3 Evidence from Design-Time Controls 5.4 Evidence from Operation-Time Controls 6 Combining Safety Evidence in the Assurance Case 7 Conclusions References A Variational Deep Synthesis Approach for Perception Validation 1 Introduction 2 Related Work 3 Concept and Overview 4 VALERIE: Computational Deep Validation 4.1 Scene Generator 4.2 Validation Parameter Space (VPS) 4.3 Computation of Synthetic Data 4.4 Sensor Simulation 4.5 Computation and Evaluation of Perceptional Functions 4.6 Controller 4.7 Computational Aspects and System Scalability 5 Evaluation Results and Discussion 6 Outlook and Conclusions References The Good and the Bad: Using Neuron Coverage as a DNN Validation Technique 1 Introduction 2 Related Works 3 Granularity and Novelty 4 Formal Analysis of Achievable Coverage 4.1 Description of the Task 4.2 Experimental Setup 4.3 Experimental Results 4.4 Discussion 5 Conclusion References Joint Optimization for DNN Model Compression and Corruption Robustness 1 Introduction 2 Related Works 3 HCRC: A Systematic Approach 3.1 Preliminaries on Semantic Segmentation 3.2 Robustness Objective 3.3 Compression Objective 3.4 HCRC Core Method 4 Experimental Setup 4.1 Datasets and Semantic Segmentation Network 4.2 Image Corruptions 4.3 Metrics 4.4 Training Framework 5 Experimental Results and Discussion 5.1 Ablation Studies 5.2 Comparison With Reference Baselines 6 Conclusions References