Artificial Intelligence Research: Third Southern African Conference, SACAIR 2022, Stellenbosch, South Africa, December 5–9, 2022, Proceedings

Description

This book constitutes the refereed proceedings of the Third Southern African Conference on  Artificial Intelligence Research, SACAIR 2022, held in Stellenbosch, South Africa, in  December 2022.  The 26 papers presented were thoroughly reviewed and selected from the 73 submissions. They are organized on the topical sections on​ algorithmic, data driven and symbolic AI; socio-technical and human-centered AI; responsible and ethical AI.  Preface Message from the General Chairs Organization Contents Algorithmic, Data Driven and Symbolic AI Adversarial Training for Channel State Information Estimation in LTE Multi-antenna Systems 1 Introduction 2 Background 2.1 Channel State Information 2.2 Super Resolution GAN 2.3 Diversity Techniques 3 Related Work 4 Experimental Setup 4.1 Dataset 4.2 System Description 4.3 Training Protocol 5 Analysis 5.1 Sample Size Selection Using ResNet 5.2 Adversarial Network Training 5.3 Receiver Diversity 5.4 Transmit Diversity 6 Conclusion References Content-Based Medical Image Retrieval Using a Class Similarity-Aware Cross-Entropy Loss 1 Introduction 2 Related Work 3 Proposed Method 3.1 IRMA Dataset 3.2 CNN-Based Classification 3.3 Evaluation Metrics 4 Experiments and Results 4.1 Experiments 4.2 Results 5 Conclusion References Jacobian Norm Regularisation and Conditioning in Neural ODEs 1 Introduction 2 Background and Definitions 3 Methodology 4 Results 4.1 Generalisation and Sensitivity 4.2 Jacobian Norms and Condition Numbers 4.3 Distance to Decision Boundary 5 Additional Datasets 6 Related Work 7 Conclusion References Improving Cause-of-Death Classification from Verbal Autopsy Reports 1 Introduction 2 Background 2.1 Verbal Autopsies 2.2 Transfer Learning 3 Methods 3.1 Algorithms 3.2 Dataset 3.3 Class Imbalance 4 Results and Discussion 5 Conclusion References Real Time In-Game Playstyle Classification Using a Hybrid Probabilistic Supervised Learning Approach 1 Introduction 2 Related Work 3 Background 3.1 Play Log Definition 3.2 Playstyle Set Definition 3.3 Game Levels 3.4 Playstyle In-Game Classification Problem Definition 4 Playstyle Classification Method 4.1 Trajectory Processing 4.2 Playstyle Classification 5 Evaluation by Experiments 5.1 Case I: MiniDungeons Experiment 5.2 Case I: MiniDungeons Experimental Results 5.3 Case II: Super Mario Bros Experiment 5.4 Case II: Super Mario Bros Experimental Results 6 Conclusion and Future Work References The Missing Margin: How Sample Corruption Affects Distance to the Boundary in ANNs 1 Introduction 2 Related Work 3 Formulating the Classification Margin 4 Experimental Setup 4.1 Controlled Noise 4.2 MNIST Models 4.3 CIFAR10 Models 4.4 Terminology 5 Results 5.1 Local Inconsistencies 5.2 Discussion 6 A Deeper Look 7 Conclusion References ST-GNNs for Weather Prediction in South Africa 1 Introduction 2 Background and Related Work 2.1 Problem Formulation 2.2 Deep Neural Networks for Weather Prediction 2.3 Low Rank Weighted Graph Neural Network (WGN) 2.4 Graph WaveNet (GWN) 3 Experimental Design 3.1 Data 3.2 Pre-processing 3.3 Walk-Forward Validation 3.4 Baseline TCN and LSTM Models 3.5 ST-GNNs 3.6 Implementation 4 Results 4.1 Results Summary 4.2 Performance at Different Weather Stations 4.3 Spatial-Temporal Dependency Analysis 4.4 Spatial-Temporal Dependencies 5 Discussion and Conclusions 6 Limitations and Future Work References Multi-modal Recommendation System with Auxiliary Information 1 Introduction 2 Background and Related Work 3 Experimental Methodology 3.1 Problem Statement 3.2 Multi-modal Auxiliary Information 3.3 Embedding Layer 3.4 Transformers 3.5 Datasets 3.6 Baselines 3.7 Evaluation 4 Results 4.1 Ablation Study 4.2 Visualizing Attention Weights 5 Conclusion References Cauchy Loss Function: Robustness Under Gaussian and Cauchy Noise 1 Introduction 2 Background 2.1 Consequences of the Gaussian Assumption 2.2 Stable Distributions 2.3 Cauchy Distribution and Cauchy Loss Function 2.4 Deterministic Noise 2.5 On the Validity of Inferring from the Results 2.6 Related Work 3 Methodology 3.1 Handcrafted Experiments 3.2 Seoul Bike Sharing Demand Experiment 3.3 General Setup 3.4 General Procedure 4 Discussion 4.1 2-Variable Handcrafted Experiment 4.2 8-Variable Handcrafted Experiment 4.3 Seoul Bike Sharing Demand Experiment 5 Conclusion References CASA: Cricket Action Similarity Assessment in Video Footage Using Deep Metric Learning 1 Introduction 2 Problem Background 2.1 Related Works 3 Experiment Setup 3.1 Methods 4 Results 4.1 Ablation Study 5 Discussion of Results 6 Conclusion References From GNNs to Sparse Transformers: Graph-Based Architectures for Multi-hop Question Answering 1 Introduction 2 Background 2.1 Message Passing GNNs 2.2 Attention 2.3 GAT 2.4 Transformer 2.5 Gating and Over-Smoothing 3 Model 3.1 Graph Construction 3.2 Graph Node Embedding 3.3 GNN Encoding 3.4 Output Model 4 Experimental Setup 4.1 Implementation 5 Results 5.1 GNN Architecture 5.2 Graph Structure and Edge Embeddings 6 Discussion 7 Conclusion References Towards a Methodology for Addressing Missingness in Datasets, with an Application to Demographic Health Datasets 1 Introduction 1.1 Problem Statement 1.2 Objectives 1.3 Contribution 2 Background 2.1 Causes of Missing Data 2.2 Categories of Missing Data 2.3 Tracking Missing Data 3 Related Work 3.1 Missing Data Imputation Methods 4 Methods 4.1 Approach 4.2 Metrics of Interest 5 Results and Discussion 5.1 Synthetic Data 5.2 Classification 5.3 Direct Analysis of Imputation 6 Conclusion References Defeasible Justification Using the KLM Framework 1 Introduction 2 Background 3 Defeasible Justification Algorithm 4 Defeasible Justification Implementation 4.1 Algorithm Implementation 4.2 Testing and Evaluation 5 Conclusion and Future Work References Relevance in the Computation of Non-monotonic Inferences 1 Introduction 2 Preliminaries 2.1 Propositional Logic 2.2 Reasoning with Nonmonotonic Conditionals 2.3 Inductive Inference 2.4 System Z 2.5 Lexicographic Entailment 2.6 Computational Complexity 3 Computational Complexity for Inductive Inference 4 Algorithms for Lexicographic Closure 5 Splitting a Conditional Knowledge Base 6 A General Result 7 Related Work 8 Conclusion and Future Work References Adaptive Reasoning: An Affect Related Feedback Approach for Enhanced E-Learning 1 Introduction 2 Related Work: Affect Analysis and Reasoning 2.1 Affect Analysis 2.2 Reasoning and Its Applications 3 Proposed Framework 3.1 Data and Feature Sampling 3.2 Training and Validating Bi-LSTM Model 3.3 Experimental Results on DAiSEE 3.4 Affective States to Learning Affects (ASLA): Initial Mappings 3.5 Affective States to Basic Emotion: Validating ASLA Mapping 3.6 Live Testing of the Proposed Model 4 Conclusion and Future Work References TransFusion: Transcribing Speech with Multinomial Diffusion 1 Introduction 2 Related Work 2.1 Connectionist Temporal Classification 2.2 Denoising Diffusion Probabilistic Models 2.3 Multinomial Diffusion 3 Model 3.1 Conditioning Diffusion on Speech Representations 3.2 Training Task 3.3 Architecture 4 Diffusion Decoding 4.1 Resampling 4.2 Sequentially Progressive Diffusion 4.3 Classifier-Free Guidance 4.4 Full Inference Process 5 Experimental Setup 5.1 Dataset and Metrics 5.2 Baseline Models 5.3 TransFusion Implementation 6 Results 7 Conclusion References Fine-Tuned Self-supervised Speech Representations for Language Diarization in Multilingual Code-Switched Speech 1 Introduction 2 Background 2.1 Language Diarization 3 Corpus 4 Models 4.1 BiLSTM 4.2 X-vector Self-Attention 4.3 WavLM 5 Experimental Procedure 5.1 Data Preparation and Feature Extraction 5.2 Evaluation Metrics 6 Results and Discussion 7 Conclusion 7.1 Limitations and Future Work References Evaluating Automated and Hybrid Neural Disambiguation for African Historical Named Entities 1 Introduction 2 Related Works 2.1 Historical NED 2.2 Low-Resource NED 2.3 South African NLP 3 Data Collection 3.1 Document Selection 3.2 Document Annotation 3.3 Fold Creation 4 Baseline 4.1 Architecture 4.2 Mention Detection 4.3 Entity Selection 5 Automatic NED System 5.1 Architecture 5.2 Training 6 Results 6.1 Evaluation 6.2 Comparison with the Baseline 6.3 Performance by Document Type 7 Hybrid NED 7.1 Mention Detection 7.2 Entity Linking 7.3 Evaluation 7.4 Comparison with Automatic NED System 8 Conclusion References Neural Speech Processing for Whale Call Detection 1 Introduction 2 Related Work 3 Approach 3.1 Neural Speech Processing 3.2 Speech Features 4 Dataset 4.1 The AADC Dataset 4.2 Data Processing and Event Selection 5 CNN Baseline 5.1 Additional Data Processing 5.2 Architecture 5.3 Optimisation Protocol 6 Whale Call Detection Using Speechbrain 6.1 Framing Whale Call Detection as Different Machine Learning Tasks 6.2 Additional Data Processing 6.3 Architecture 6.4 Optimisation Protocol 7 Analysis and Results 7.1 CNN 7.2 Speechbrain's TDNN 8 Conclusion References Self-Supervised Text Style Transfer with Rationale Prediction and Pretrained Transformers 1 Introduction 2 Background 3 Sentiment Word Identification 3.1 Saliency Noising 3.2 Rationales Noising 4 Self-supervised Text Sentiment Transfer 5 Experimental Setup 5.1 Data 5.2 Sentiment Word Identification 5.3 Sentiment Transfer Models 6 Results 6.1 Automatic Evaluation 6.2 Human Evaluation 6.3 Qualitative Analysis 7 Conclusion References Socio-Technical and Human-Centered AI AI for Social Good: Sentiment Analysis to Detect Social Challenges in South Africa 1 Introduction 2 Related Work 3 Experimental Setup 3.1 Overview of Our Systems 3.2 Collection of SAGovTopicTweets 3.3 The SAfriSenti Corpus 3.4 Sentiment Analysis 4 Experiments and Results 4.1 Language Distributions of Collected Tweets 4.2 Sentiment Analysis to Detect Social Challenges 5 Conclusion and Future Work References A Model for Biometric Selection in Public Services Sector 1 Introduction 2 Problem Background 3 Related Work on Technology Acceptance 4 Evaluation of Biometric Authentication Methods 5 A Model Discussion for Multimodal Biometric Selection 5.1 Usability 5.2 Public Awareness 5.3 Public Perception and Acceptability 5.4 Public Satisfaction 5.5 Confidentiality, Integrity, Availability and Learnability (CIAL) 6 Conclusion References Technology Days: An AI Democratisation Journey Begins with a Single Step 1 Introduction 2 Background 2.1 Democratisation 2.2 Democratisation of AI in the Context of a Developing Country 2.3 Democratisation and Technology Acceptance 3 Case Study 3.1 The Study Setting 3.2 Data Collection 3.3 Data Analysis 4 Discussion 5 Conclusion References The Preparation of South African Companies for the Impact of Artificial Intelligence 1 Introduction 2 Background and Related Work 3 Research Approach 4 Results 4.1 Demographics 4.2 Understanding AI 4.3 Preparing for AI 4.4 Expected Impact of AI 4.5 Alignment of Strategy 5 Discussion 6 Conclusion Appendix: Survey Questions References Responsible and Ethical AI Answerability, Accountability, and the Demands of Responsibility 1 Responsibility and AI 2 Responsibility Gaps and AI 3 Types of Responsibility 4 Responsibility as Answerability 4.1 The Wrong Kind of Reason 4.2 Collaborative AI 4.3 The “Problem” of Many Hands 5 Responsibility as Accountability 5.1 Retribution Gaps 6 Conclusion References Does Counterfactual Reasoning Hold the Key to Artificial General Intelligence? 1 Introduction 2 A Critique of Judea Pearl and Dana Mackenzie 2.1 Pearl and Mackenzie’s Mini-Turing Test 2.2 Pearl and Mackenzie’s Conflations 3 The Counterfactual Room Argument 3.1 Premise One: Consciousness is a Necessary Condition for Imagination 3.2 Premise 2: Imagination is a Necessary Condition for the Ability to Reason Counterfactually 3.3 Premise 3: The Ability to Reason Counterfactually is a Necessary Condition for an Agent to Demonstrate Understanding 3.4 Premise 4: Understanding is a Necessary Condition for an Agent to Pass a Requisite Version of the Turing Test for AGI 3.5 Final Conclusion: Therefore, Consciousness is a Necessary Condition for an Agent to Pass a Requisite Version of the Turing Test for AGI 4 Conclusion References Author Index