KI 2021: Advances in Artificial Intelligence: 44th German Conference on AI, Virtual Event, September 27 – October 1, 2021, Proceedings (Lecture Notes in Computer Science)

This book constitutes the refereed proceedings of the 44 th German Conference on Artificial Intelligence, KI 2021, held in September/October 2021. Due to COVID-19 pandemic the conference was held virtually. The 16 full and 4 short papers with one extended abstract were carefully reviewed and selected from 59 submissions. As well-established annual conference series KI is dedicated to research on theory and applications across all methods and topic areas of AI research. Preface Special Events Tutorial Workshops Organization Abstracts of Invited Talks Monte Carlo Search Autonomy in AI: Reactive Synthesis, Planning and Reinforcement Learning in Linear Temporal Logic on Finite Traces Ontologies for Providing Map Knowledge to Autonomous Vehicles The Third Wave of AI Motion Intelligence for Human-Centred Robots Human-Compatible Artificial Intelligence Contents Technical Programme RP-DQN: An Application of Q-Learning to Vehicle Routing Problems 1 Introduction 2 Related Work 3 Problem Definition 4 Method 4.1 Original Attention-Model 4.2 RP-DQN 5 Experiments 5.1 Baselines 5.2 Data 5.3 CVRP Results 5.4 MDVRP Results 5.5 Learning Curves 5.6 Runtime Comparison 5.7 Generalization Study 6 Conclusion References -Circulant Maximum Variance Bases 1 Introduction 2 Preliminaries 2.1 Principal Component Analysis 2.2 Dynamic Principal Component Analysis 2.3 -Circulant Matrices 3 Maximum Variance Bases 3.1 Simple Matched Circulants 3.2 Matched -Circulant Matrices 3.3 Relation to PCA, DPCA and DFT 4 Numerical Results 4.1 MA Process 4.2 Circular Process 5 Conclusion References Quantified Boolean Solving for Achievement Games 1 Introduction 2 Quantified Boolean Formulas 3 Harary's Tic-Tac-Toe 4 Related Work 5 The Pairing Encoding 6 Experimental Results 7 Conclusion References Knowledge Graph Based Question Answering System for Financial Securities 1 Introduction 2 Framework 2.1 Knowledge Graph Construction 2.2 Semantic Question Answering System 3 Experiments 4 Conclusion References Semi-unsupervised Learning: An In-depth Parameter Analysis 1 Introduction 2 Related Work 3 Methodology 3.1 Learning Paradigms 3.2 Semi-unsupervised Learning with Deep Generative Models 4 Datasets 5 Experiments 5.1 Semi-unsupervised Classification 5.2 Parameter Analysis 6 Summary and Outlook References Combining Transformer Generators with Convolutional Discriminators 1 Introduction 2 Related Work 2.1 Generative Models Using CNNs 2.2 Generative Models Using Attention 2.3 Hybrid Models 3 Model Architecture 3.1 Transformer Generator 3.2 Convolutional Discriminator 4 Experiments 4.1 Setup 4.2 Results 4.3 Frequency Analysis 5 Discussion 6 Conclusion References Explanation as a Process: User-Centric Construction of Multi-level and Multi-modal Explanations 1 Introduction 2 A Relational Knowledge Domain 3 Learning an Interpretable Model with ILP 4 Multi-level and Multi-modal Explanations 4.1 Explanation Generation 4.2 Explanatory Dialogue 4.3 Proof-of-Concept Implementation 5 Conclusion and Outlook References Multi-Type-TD-TSR – Extracting Tables from Document Images Using a Multi-stage Pipeline for Table Detection and Table Structure Recognition: From OCR to Structured Table Representations 1 Introduction 2 Related Work 3 End-to-End Multistage Pipeline 4 Methods 4.1 Table Alignment Pre-processing 4.2 Table Detection 4.3 Bordered TSR 4.4 Unbordered TSR 4.5 Partially Bordered TSR 4.6 Color Invariance Pre-Processing 5 Evaluation 6 Conclusion References A High-Speed Neural Architecture Search Considering the Number of Weights 1 Introduction and Related Works 2 Proposed Method 2.1 DARTS Algorithm 2.2 Loss Function with the Number of Weights 3 Experiments 3.1 Implementation Details 3.2 Results 4 Conclusions References Semantic Segmentation of Aerial Images Using Binary Space Partitioning 1 Introduction 2 Related Work 3 Semantic Segmentation 3.1 BSP-based Segmentation Model 3.2 Differentiable BSP Tree Rendering 3.3 Comparison to State of the Art 4 Evaluation 4.1 Datasets 4.2 Training and Test Setup 4.3 Ground Truth as BSP Trees 4.4 Semantic Segmentation 4.5 Prediction Confidence 5 Conclusion A Dataset Class Distributions B Hyperparameters and Model Details C Metrics D Additional Sample Images E Confidence References EVARS-GPR: EVent-Triggered Augmented Refitting of Gaussian Process Regression for Seasonal Data 1 Introduction 2 Related Work 3 Problem Formulation 4 EVARS-GPR 5 Experimental Setup 5.1 Simulated Data 5.2 Real-World Datasets 5.3 Evaluation 6 Experimental Results 6.1 Behavior on Simulated Data 6.2 Results on Real-World Datasets 6.3 Discussion 7 Conclusion Appendix A: Gaussian Process Regression Appendix B: List of Symbols Appendix C: Online Change Point Detection Appendix D: Data Augmentation Appendix E: EVARS-GPR Parameters Appendix F: Real-World Datasets Appendix G: Further Simulated Scenarios References Selective Pseudo-Label Clustering 1 Introduction 2 Related Work 3 Method 3.1 Formal Description 3.2 Implementation Details 4 Proof of Correctness 4.1 Agreed Pseudo-Labels Are More Accurate 4.2 Increased Pseudo-Label Accuracy Improves Clustering 5 Experimental Results 5.1 Ablation Studies 5.2 Ensemble Size 6 Conclusion A Appendix A: Full Proofs A.1 More Accurate Pseudo-Labels Supplement A.2 Lemma 1 Supplement A.3 Lemma 2 Supplement A.4 Lemma 3 Supplement A.5 Theorem 5 Supplement B Appendix C: Extended Results References Crop It, but Not Too Much: The Effects of Masking on the Classification of Melanoma Images 1 Introduction 2 Related Work 3 Method Overview 4 Experiments 5 Visual Inspection 6 Discussion 7 Conclusions A MedNode Results B ISIC2016 Results References A Demonstrator for Interactive Image Clustering and Fine-Tuning Neural Networks in Virtual Reality 1 Introduction 2 Architecture 3 Visualization in VR 3.1 PCA/t-SNE Approach 3.2 VAE Approach 4 Interactive Fine-Tuning 5 Conclusion and Future Work References HUI-Audio-Corpus-German: A High Quality TTS Dataset 1 Introduction 2 Related Work 3 Data Processing Pipeline 3.1 Acquisition of Suitable Audio Data 3.2 Splitting of Audio Data 3.3 Audio Normalization 3.4 Transcription of Audio Data for Subsequent Alignment 3.5 Acquisition of Text for Audio Data 3.6 Text Normalization 3.7 Transcript Alignment 4 Dataset Summary 4.1 Full Dataset 4.2 Clean Dataset 4.3 Discussion 4.4 Evaluation with Tacotron 2 5 Conclusion and Outlook References Negation in Cognitive Reasoning 1 Introduction 2 Background and Related Works 2.1 Negation in Logic and Natural Language 2.2 Commonsense Reasoning and Negation 3 Methods 3.1 A System for Cognitive Reasoning 3.2 Negation Scope and the Negatus – Why Size Matters 3.3 Approach to Negation Treatment for Cognitive Reasoning 4 Experiments 4.1 Data Preparation and Evaluation 5 Summary, Conclusions, and Future Work References Learning to Detect Adversarial Examples Based on Class Scores 1 Introduction 1.1 Related Work 1.2 Contributions 2 Detecting Adversarial Examples from Class Scores 3 Experimental Setup 4 Results 5 Conclusion References An Agent Architecture for Knowledge Discovery and Evolution 1 Introduction 2 Background and Related Work 2.1 The BDI Architecture 2.2 Integrating AI into BDI Agents 2.3 KDE Systems and Approaches 3 Design of the KDE Agent Architecture 4 The KDE Agent Architecture 4.1 The Exogenous Modules 4.2 The Endogenous Modules 5 Use Case - Domestic Electricity Consumption 5.1 Cluster Analysis Service 5.2 Perception 5.3 Deliberation to Generate Explanations 6 Discussion and Conclusion References Demystifying Artificial Intelligence for End-Users: Findings from a Participatory Machine Learning Show 1 Introduction 2 Related Work 2.1 Virtual Agents in Education and Edutainment 2.2 Explainable AI 2.3 Trust in Technical Systems 3 Field Study 3.1 Demonstrator Setup 3.2 Study Procedure 3.3 Evaluation Method 4 Results 4.1 Information About Participants 4.2 Results of the ML-show 4.3 Comparison Between Participating and Non-participating Museum Visitors 5 Discussion 5.1 Take Users' Attitudes and Experiences into Account 5.2 Think About Who You Want to Reach with XAI Edutainment 5.3 Trust and Distrust Are Important Components in XAI Interaction Design 6 Conclusion References Recent Advances in Counting and Sampling Markov Equivalent DAGs 1 Introduction 2 Main Results References An Approach to Reduce the Number of Conditional Independence Tests in the PC Algorithm 1 Introduction 2 Preliminaries 3 Detection of V-Structures in Advance 4 The ED-PC Algorithm 5 Proof of Correctness 6 Experimental Analysis 7 Conclusions and Outlook References Poster Papers Unsupervised Anomaly Detection for Financial Auditing with Model-Agnostic Explanations 1 Introduction 2 Related Work 3 Explainable Anomaly Detection in the Context of Auditing 3.1 Data 3.2 Feature Engineering 3.3 Ensemble-Based Architecture 3.4 Model-Agnostic and Receiver-Dependent Explanations 4 Conclusion and Future Work A Anomaly Detection Ensemble B Explanations References A Priori Approximation of Symmetries in Dynamic Probabilistic Relational Models 1 Introduction 2 Preliminaries 3 A Priori Approximation of Symmetries 3.1 Encoding Model Behavior by Ordinal Pattern Symbolization 3.2 Entity Symmetry Approximation by Spectral Clustering 4 Preventing Groundings 4.1 Avoiding Model Splits with Evidence 4.2 Interconnectivity Yielding Inferred Evidence 5 Empirical Evaluation 6 Conclusion and Future Work References Towards More Explicit Interaction Modelling in Agent-Based Simulation Using Affordance Schemata 1 Motivation 2 Affordances in Modelling 3 Agent-Entity Interaction 3.1 Phases in the Behaviour of an Agent, Agent and Activities 3.2 The Interaction Perspective 3.3 Affordances as Guidance for Interactions 4 Affordance Schemata as Framework for Modelling Agent-Entity Interactions 4.1 Affordance and Affordance Schema 4.2 Decision Making About Interactions Based on Affordances 5 Proof of Concept 5.1 Conceptual Example 5.2 Implementation 6 Discussion: Interaction as First-Class Concept 7 Related Work 8 Summary and Future Work References Critic Guided Segmentation of Rewarding Objects in First-Person Views 1 Introduction 2 Methods 2.1 Pipeline 3 Experiments 4 Results 5 Discussion References Self-supervised Domain Adaptation for Diabetic Retinopathy Grading Using Vessel Image Reconstruction 1 Introduction 2 Related Work 3 Method 3.1 Overview 3.2 Retinal Vessel Reconstruction-Based SSL 3.3 Relevant Features from SSL 4 Experiments and Results 4.1 Evaluation Method 4.2 Dataset and Metrics 4.3 Performance of Unsupervised Adaption Methods 4.4 Performance of Baseline Methods on DR Grading Prediction 5 Conclusion References BehavE: Behaviour Understanding Through Automated Generation of Situation Models 1 Introduction 2 Related Work 3 Automatic Generation of Situation Models 3.1 Knowledge Extraction and Consolidation 3.2 Generation of Computational Causal Behaviour Models 4 Planned Experiments 5 Conclusion and Future Work References Author Index