Artificial Intelligence and Machine Learning: 33rd Benelux Conference on Artificial Intelligence, BNAIC/Benelearn 2021, Esch-sur-Alzette, Luxembourg, November ... Computer and Information Science Book 1530)

This book contains a selection of the best papers of the 33rd Benelux Conference on Artificial Intelligence, BNAIC/ BENELEARN 2021, held in Esch-sur-Alzette, Luxembourg, in November 2021. The 14 papers presented in this volume were carefully reviewed and selected from 46 regular submissions. They address various aspects of artificial intelligence such as natural language processing, agent technology, game theory, problem solving, machine learning, human-agent interaction, AI and education, and data analysis. Preface Organization Contents Annotating Data Active Learning for Reducing Labeling Effort in Text Classification Tasks 1 Introduction 2 Related Work 3 Methods 3.1 Active Learning 3.2 Model Architecture 3.3 Query Functions 3.4 Heuristics 3.5 Experimental Setup 4 Results 4.1 Active Learning 4.2 Query-Pool Size 4.3 Heuristics 5 Discussion A.1 RET Algorithm Computational Cost Analysis A.2 Algorithms References Refining Weakly-Supervised Free Space Estimation Through Data Augmentation and Recursive Training 1 Introduction 2 Related Work 2.1 Supervised Learning for Segmentation 2.2 Weakly-Supervised Semantic Segmentation 2.3 Unsupervised and Weakly-Supervised Monocular Free Space Segmentation 2.4 Training Strategies for Weakly-Supervised Segmentation 3 Methodology 3.1 Data Augmentation 3.2 Recursive Training 4 Experimental Setup 4.1 Dataset 4.2 Evaluation Metrics 4.3 Network Architectures 4.4 Training Procedure 4.5 Use of Ground Truth Data 5 Results 5.1 Fully-Supervised Results 5.2 Unsupervised and Weakly-Supervised Baselines 5.3 Data Augmentation and Recursive Training 5.4 Limits of Recursive Training 5.5 Qualitative Results 6 Conclusion References Self-labeling of Fully Mediating Representations by Graph Alignment 1 Introduction 2 Related Work 3 Self-labeling of Fully Mediating Representations 3.1 Graph Alignment 3.2 Method 4 Experiments 5 Conclusion A Appendix A.1 Architecture Summary of Graph Recognition Tool A.2 Training Details for Graph Recognition Tool A.3 Computational Cost per Rich-Labeling Iteration A.4 Examples of Cases Where Graph Alignment Fails References Recognizing Objects Task Independent Capsule-Based Agents for Deep Q-Learning 1 Introduction 2 Background 2.1 Capsule Networks 2.2 Deep Reinforcement Learning 3 Related Work 4 Methodology 5 Analysis 5.1 Cumulative Reward and Parameters 5.2 Input State 5.3 Action Space 5.4 Reward 6 Discussion 6.1 Training 6.2 Environment 7 Conclusion References Object Detection with Semi-supervised Adversarial Domain Adaptation for Real-Time Edge Devices 1 Introduction 2 Related Work 2.1 Object Detection 2.2 Domain Adaptation 3 Proposed Method 3.1 Adversarial Domain Adaptation for Object Detection 4 Implementation Details 5 Evaluation 5.1 Datasets 5.2 Experiments 6 Conclusion References Explaining Outcomes Exploring Explainable AI in the Financial Sector: Perspectives of Banks and Supervisory Authorities Abstract 1 Introduction 2 Theoretical Background 3 Research Method 3.1 Use Cases 3.2 Data Collection 3.3 Data Analysis 4 Results 4.1 Consumer Credit 4.2 Credit Risk Management 4.3 Anti-money Laundering (AML) 4.4 General 5 Discussion and Conclusions References The Effect of Noise Level on the Accuracy of Causal Discovery Methods with Additive Noise Models 1 Introduction 2 Related Work 3 Causal Discovery Methods 3.1 Notations 3.2 Regression with Subsequent Independence Test (Resit) 3.3 Identification Using Conditional Variances (Uncertainty Scoring) 4 Experimental Setup 5 Experimental Results 5.1 Resit 5.2 Uncertainty Scoring 6 Conclusions References A Bayesian Framework for Evaluating Evolutionary Art 1 Introduction 2 Background 2.1 Evaluating Computer-Generated Art 3 The Bayesian Framework 3.1 Art Turing Test 3.2 Bayesian Model Comparison 4 Application 4.1 Tree Representation 4.2 The Mathematical Fitness Function 4.3 Results and Analysis 5 Questionnaire 6 Code Base 7 Discussion 8 Conclusion References Understanding Language Dutch SQuAD and Ensemble Learning for Question Answering from Labour Agreements 1 Introduction 2 Related Work 3 Datasets 3.1 Dutch SQuAD v2.0 3.2 Labour Agreement Dataset 4 Approach 4.1 Fine-Tuning 4.2 Voted BERT 5 Evaluation 5.1 Models 5.2 Evaluation Metrics 6 Results 6.1 Dutch SQuAD 6.2 Labour Agreement Dataset 7 Discussion 7.1 Conclusion References Verbalizing but Not Just Verbatim Translations of Ontology Axioms 1 Introduction 2 Related Work 3 Preliminaries and Defintions 4 Proposed Verbalization Approach 5 Semantic-Refinement of Label-Sets 6 Empirical Evaluation 6.1 Results and Discussions 7 Conclusion References Transfer Learning and Curriculum Learning in Sokoban 1 Introduction 2 Related Work 3 Experimental Setup 3.1 Neural Network Architecture 3.2 Transfer Approach 4 Experiments 4.1 Transfer Among Related Tasks 4.2 Transfer Among Different Tasks (SL/RL) 4.3 Transfer to Different Appearance 4.4 Visualizing Agent Detection 5 Conclusion and Future Work References Reinforcing Decisions Proximal Policy Optimisation for a Private Equity Recommitment System 1 Introduction 2 Related Works 3 Problem Description 4 Proximal Policy Optimisation 5 Private Equity Recommitment as RL Problem 5.1 Modelling 5.2 Synthetic Cashflows 6 Experimental Setups 7 Experimental Results 8 Conclusion References Regular Decision Processes for Grid Worlds 1 Introduction 2 Background 2.1 Markov Decision Processes 2.2 Non-Markovian Decision Processes 2.3 Temporal Logic, Automata and Product MDPs 2.4 Regular Decision Processes: Non-Markovian Dynamics 2.5 Related Work 3 Approach and Software Design 3.1 Compilation: From RDP to MDP 4 Experiments 4.1 Experiment 1: Goal Sparsity 4.2 Experiment 2: Reward Shaping 4.3 Experiment 3: Safety 4.4 Experiment 4: Non-Markovian Transitions 5 Conclusions and Future Work References MoveRL: To a Safer Robotic Reinforcement Learning Environment 1 Introduction 2 Notations 3 Related Work 3.1 RL Robotic Environment 3.2 Safe Reinforcement Learning 3.3 Path Planning 4 Contribution 4.1 The Gym Environment 4.2 Observation Space 4.3 Action Space 4.4 Why Do We Need Sequences of Actions? 4.5 Reward Function 4.6 Initial States and Termination 4.7 Safety Guarantee 5 Experiment 5.1 Learning Scenarios 5.2 Learning Algorithm 5.3 Results 6 Conclusion References Author Index Explainable artificial intelligence (xAI) is seen as a solution to making AI systems less of a “black box”. It is essential to ensure transparency, fairness, and accountability - which are especially paramount in the financial sector. The aim of this study was a preliminary investigation of the perspectives of supervisory authorities and regulated entities regarding the application of xAI in the financial sector. Three use cases (consumer credit, credit risk, and anti-money laundering) were examined using semi-structured interviews at three banks and two supervisory authorities in the Netherlands. We found that for the investigated use cases a disparity exists between supervisory authorities and banks regarding the desired scope of explainability of AI systems. We argue that the financial sector could benefit from clear differentiation between technical AI (model) explainability requirements and explainability requirements of the broader AI system in relation to applicable laws and regulations