Deep Learning Theory and Applications: Third International Conference, DeLTA 2022, Lisbon, Portugal, July 12–14, 2022, Revised Selected Papers (Communications in Computer and Information Science)

This book constitutes the refereed post-conference proceedings of the Third International Conference on Deep Learning Theory and Applications, DeLTA 2022, held in Lisbon, Portugal, during January 17-18, 2022. The 6 full papers included in this book were carefully reviewed and selected from 36 submissions. They present recent research on machine learning and artificial intelligence in real-world applications such as computer vision, information retrieval and summarization from structured and unstructured multimodal data sources, natural language understanding and translation, and many other application domains. Preface Organization Contents Modified SkipGram Negative Sampling Model for Faster Convergence of Graph Embedding 1 Introduction 2 Related Work 3 Graph Embedding: Random Walk Based Technique 3.1 Standard SkipGram Model: Negative Sampling Approach 3.2 Proposed SkipGram Model 3.3 Community Detection and Link Prediction Tasks 4 Experimental Evaluation 4.1 Data Description and Implementation Details 4.2 Evaluation of DeepWalkb Model 5 Conclusions References Active Collection of Well-Being and Health Data in Mobile Devices 1 Introduction 2 Models for Smart Notification and Users 2.1 Reinforcement Learning Algorithms 2.2 State Representation 2.3 Reward Definition 2.4 Environment Model - Simulation 3 Experimental Description and Metrics 3.1 Model Initialization Methods 3.2 Daily vs. Weekly Routine 3.3 Performance Metrics 3.4 Implementation of the Smart Notification App 4 Results and Discussion 4.1 Simulation Results 4.2 Real-World Pilot 5 Conclusions and Future Work References Reliable Classification of Images by Calculating Their Credibility Using a Layer-Wise Activation Cluster Analysis of CNNs 1 Introduction 2 Related Work 3 Method 3.1 Identifying Clusters in Layer Activations 3.2 Obtaining In-Distribution Statistics 3.3 Calculating the Credibility of an Image at Inference 4 Experiments 4.1 Baseline 4.2 Testing with Fewer Layers 4.3 Testing on ImageNet Data 5 Conclusion References Traffic Sign Repositories: Bridging the Gap Between Real and Synthetic Data 1 Introduction 2 Synthetic vs. Real Traffic Sign Datasets 3 Related Work 3.1 Traffic Sign Datasets 3.2 Synthetic Traffic Signs 4 Synthetic Traffic Signs Generation Algorithm 4.1 Background 4.2 Brightness Distribution 4.3 Confetti Noise 4.4 Perlin Noise 4.5 Synthetic Generation Pipeline 5 Evaluation 5.1 Solo Synthetic Dataset Evaluation 5.2 Combining Real and Synthetic Data 5.3 Cross-Testing 5.4 Unleashing Synthetic Datasets 6 Conclusion References Convolutional Neural Networks for Structural Damage Localization on Digital Twins 1 Introduction and Background 2 Methodology and Methods 2.1 Digital Twin Design 2.2 Deep Learning for the Damage Localization 3 Case Study 4 Experiments and Results 4.1 Damage Localization via CNN 4.2 Noise Tolerance Evaluation 4.3 Random Search Algorithm for Hyperparameter Optimization 5 Conclusion References Evaluating and Improving RoSELS for Road Surface Extraction from 3D Automotive LiDAR Point Cloud Sequences*-4pt 1 Introduction 2 Related Work 2.1 Road Edge Detection 2.2 Scene Classification 2.3 Ground Plane Estimation 2.4 Surface Reconstruction 3 Background: RoSELS System Description 3.1 Ground Point Detection (S1) 3.2 Road Edge Point Detection (S2a) 3.3 Frame Classification (S2b) 3.4 Edge Point Set Smoothing (S3) 3.5 3D Road Surface Extraction (S4) 4 Evaluating and Improving RoSELS 5 Experiments and Results 5.1 Implementation of Road Surface Extraction 5.2 Experiments and Results 6 Conclusions References Author Index