Meta-Learning : Theory, Algorithms and Applications

Deep neural networks (DNNs) with their dense and complex algorithms provide real possibilities for Artificial General Intelligence (AGI). Meta-learning with DNNs brings AGI much closer: artificial agents solving intelligent tasks that human beings can achieve, even transcending what they can achieve. Meta-Learning: Theory, Algorithms and Applications shows how meta-learning in combination with DNNs advances towards AGI. Meta-Learning: Theory, Algorithms and Applications explains the fundamentals of meta-learning by providing answers to these questions: What is meta-learning?; why do we need meta-learning?; how are self-improved meta-learning mechanisms heading for AGI?; how can we use meta-learning in our approach to specific scenarios? The book presents the background of seven mainstream paradigms: meta-learning, few-shot learning, deep learning, transfer learning, machine learning, probabilistic modeling, and Bayesian inference. It then explains important state-of-the-art mechanisms and their variants for meta-learning, including memory-augmented neural networks, meta-networks, convolutional Siamese neural networks, matching networks, prototypical networks, relation networks, LSTM meta-learning, model-agnostic meta-learning, and the Reptile algorithm. The book takes a deep dive into nearly 200 state-of-the-art meta-learning algorithms from top tier conferences (e.g. NeurIPS, ICML, CVPR, ACL, ICLR, KDD). It systematically investigates 39 categories of tasks from 11 real-world application fields: Computer Vision, Natural Language Processing, Meta-Reinforcement Learning, Healthcare, Finance and Economy, Construction Materials, Graphic Neural Networks, Program Synthesis, Smart City, Recommended Systems, and Climate Science. Each application field concludes by looking at future trends or by giving a summary of available resources. Meta-Learning: Theory, Algorithms and Applications is a great resource to understand the principles of meta-learning and to learn state-of-the-art meta-learning algorithms, giving the student, researcher and industry professional the ability to apply meta-learning for various novel applications. A comprehensive overview of state-of-the-art meta-learning techniques and methods associated with deep neural networks together with a broad range of application areas Coverage of nearly 200 state-of-the-art meta-learning algorithms, which are promoted by premier global AI conferences and journals, and 300 to 450 pieces of key research Systematic and detailed exploration of the most crucial state-of-the-art meta-learning algorithm mechanisms: model-based, metric-based, and optimization-based Provides solutions to the limitations of using deep learning and/or machine learning methods, particularly with small sample sizes and unlabeled data Gives an understanding of how meta-learning acts as a stepping stone to Artificial General Intelligence in 39 categories of tasks from 11 real-world application fields Front Cover Meta-Learning: Theory, Algorithms and Applications Copyright Dedication Contents Preface Acknowledgments Chapter 1: Meta-learning basics and background 1.1. Introduction 1.2. Meta-learning 1.2.1. Definitions 1.2.2. Evaluation 1.2.3. Datasets and benchmarks 1.3. Machine learning 1.3.1. Models 1.3.2. Limitations 1.3.3. Related concepts 1.3.4. Further Reading 1.4. Deep learning 1.4.1. Models 1.4.2. Limitations 1.4.3. Further readings 1.5. Transfer learning 1.5.1. Multitask learning 1.6. Few-shot learning 1.7. Probabilistic modeling 1.8. Bayesian inference References Part I: Theory & mechanisms Chapter 2: Model-based meta-learning approaches 2.1. Introduction 2.2. Memory-augmented neural networks 2.2.1. Background knowledge 2.2.2. Methodology Task setup Memory retrieval Least recently used access 2.2.3. Extended algorithm 1 2.2.4. Extended algorithm 2 2.3. Meta-networks 2.3.1. Background knowledge 2.3.2. Methodology Slow weights and fast weights Layer augmentation 2.3.3. Main loss functions and representation loss functions 2.4. Summary References Chapter 3: Metric-based meta-learning approaches 3.1. Introduction 3.2. Convolutional Siamese neural networks 3.2.1. Background knowledge 3.2.2. Methodology Combination of the twin Siamese networks Objective function Optimization 3.2.3. Extended algorithm 1 3.3. Matching networks 3.3.1. Background knowledge 3.3.2. Methodology The attention kernel Full context embedding Episode-based training 3.3.3. Extended algorithm 1 3.4. Prototypical networks 3.4.1. Background knowledge 3.4.2. Methodology Bregman divergence requirement 3.4.3. Extended algorithm 1 3.4.4. Extended algorithm 2 3.4.5. Extended algorithm 3 3.5. Relation network 3.5.1. Background knowledge 3.5.2. Methodology C-Way one-shot C-Way K-shot C-Way zero-shot Objective function 3.6. Summary References Chapter 4: Optimization-based meta-learning approaches 4.1. Introduction 4.2. LSTM meta-learner 4.2.1. Background knowledge Covariate shift Batch normalization Long short-term memory Gradient-based optimization 4.2.2. Methodology Gradient independent assumption and initialization Meta-training and meta-testing batch normalization Parameter sharing 4.3. Model-agnostic meta-learning 4.3.1. Background knowledge Transfer learning Fine-tuning 4.3.2. Methodology Task adaptation 4.3.3. Illustration 1: Few-shot regression and few-shot classification 4.3.4. Illustration 2: Policy gradient reinforcement learning 4.3.5. Illustration 3: Meta-imitation learning 4.3.6. Related Algorithm 1: Meta-SGD 4.3.7. Related Algorithm 2: Feature reuse-The effectiveness of MAML 4.3.8. Related Algorithm 3: Adaptive hyperparameter generation for fast adaptation 4.4. Reptile 4.4.1. Background knowledge First-order model-agnostic meta-learning 4.4.2. Methodology 4.4.2.1. Serial version 4.4.2.2. Parallel or batch version The optimization assumption Analysis 4.4.3. Related Algorithm 1 4.4.4. Related Algorithm 2 4.4.5. Related Algorithm 3 4.4.6. Related Algorithm 4 4.5. Summary References Part II: Applications Chapter 5: Meta-learning for computer vision 5.1. Introduction 5.1.1. Limitations 5.2. Image classification 5.2.1. Introduction Development Approaches Benchmarks One-stage semisupervised learning One-stage unsupervised learning Multistage semisupervised learning 5.2.2. Decision boundary sharpness and few-shot image classification 5.2.3. Semisupervised few-shot image classification with refined prototypical network 5.2.4. Few-shot unsupervised image classification 5.2.5. One-shot image deformation 5.2.6. Heterogeneous multitask learning in image classification 5.2.7. Few-shot classification with transductive inference 5.2.8. Closed-form base learners 5.2.9. Long-tailed image classification 5.2.10. Image classification via incremental learning without forgetting Comparison and contrast of iTAML and reptile Lower bound of sample 5.2.11. Few-shot open set recognition 5.2.12. Deficiency of pretrained knowledge in few-shot learning 5.2.13. Bayesian strategy with deep kernel for regression and cross-domain image classification in a few-shot setting 5.2.14. Statistical diversity in personalized models of federated learning 5.2.15. Meta-learning deficiency in few-shot learning 5.3. Face recognition and face presentation attack 5.3.1. Introduction Facial recognition Face antispoofing 5.3.2. Person-specific talking head generation for unseen people and portrait painting in few-shot regimes 5.3.3. Face presentation attack and domain generalization 5.3.4. Anti-face-spoofing in few-shot and zero-shot scenarios 5.3.5. Generalized face recognition in the unseen domain 5.4. Object detection 5.4.1. Introduction Approaches Benchmarks 5.4.2. Long-tailed data object detection in few-shot scenarios 5.4.3. Object detection in few-shot scenarios 5.4.4. Unseen object detection and viewpoint estimation in low-data settings 5.5. Fine-grained image recognition 5.5.1. Introduction Approaches Benchmarks 5.5.2. Fine-grained visual categorization 5.5.3. One-shot fine-grained visual recognition 5.5.4. Few-shot fine-grained image recognition 5.6. Image segmentation 5.6.1. Introduction Modern development 5.6.2. Multiobject few-shot semantic segmentation 5.6.3. Few-shot static object instance-level detection 5.7. Object tracking 5.7.1. Introduction 5.7.2. Offline object tracking 5.7.3. Real-time online object tracking 5.7.4. Real-time object tracking with channel pruning One-shot channel pruning 5.7.5. Object tracking via instance detection 5.8. Label noise 5.8.1. Introduction Approaches Benchmarks 5.8.2. Reweighting examples through online approximation 5.8.3. Hallucinated clean representation for noisy-labeled visual recognition 5.8.4. Data valuation using reinforcement learning 5.8.5. Teacher-student networks for image classification on noisy labels 5.8.6. Sample reweighting function construction 5.8.7. Loss correction approach 5.8.8. Meta-relabeling through data coefficients 5.8.9. Meta-label correction 5.9. Superresolution 5.9.1. Introduction Approaches Datasets and benchmarks 5.9.2. Meta-transfer learning for zero-shot superresolution 5.9.3. LR-HR image pair superresolution 5.9.4. No-reference image quality assessment 5.10. Multimodal learning 5.10.1. Introduction Deep learning approaches Benchmarks 5.10.2. Visual question answering system 5.11. Other emerging topics 5.11.1. Domain generalization 5.11.2. High-accuracy 3D appearance-based gaze estimation in few-shot regimes 5.11.3. Benchmark of cross-domain few-shot learning in vision tasks 5.11.4. Latent embedding optimization in low-dimensional space 5.11.5. Image captioning 5.11.6. Memorization issue 5.11.7. Meta-pseudo label 5.12. Summary References Chapter 6: Meta-learning for natural language processing 6.1. Introduction 6.1.1. Limitations 6.2. Semantic parsing 6.2.1. Introduction Development Benchmarks 6.2.2. Natural language to structured query generation in few-shot learning Implementation 6.2.3. Semantic parsing in low-resource scenarios 6.2.4. Context-dependent semantic parser with few-shot learning 6.3. Machine translation 6.3.1. Introduction 6.3.2. Multidomain neural machine translation in low-resource scenarios 6.3.3. Multilingual neural machine translation in few-shot scenarios 6.4. Dialogue system 6.4.1. Introduction 6.4.2. Few-shot personalizing dialogue generation 6.4.3. Domain adaptation in a dialogue system 6.4.4. Natural language generation by few-shot learning concerning task-oriented dialogue systems 6.5. Knowledge graph 6.5.1. Introduction 6.5.2. Multihop knowledge graph reasoning in few-shot scenarios 6.5.3. Knowledge graphs link prediction in few-shot scenarios 6.5.4. Knowledge base complex question answering 6.5.5. Named-entity recognition in cross-lingual scenarios 6.6. Relation extraction 6.6.1. Introduction 6.6.2. Few-shot supervised relation classification 6.6.3. Relation extraction with few-shot and zero-shot learning 6.7. Sentiment analysis 6.7.1. Introduction Benchmark and dataset 6.7.2. Text emotion distribution learning with small samples 6.8. Emerging topics 6.8.1. Domain-specific word embedding under lifelong learning setting Background knowledge Methodology 6.8.2. Multilabel classification Background knowledge Methodology 6.8.3. Representation under a low-resource setting Background knowledge Methodology 6.8.4. Compositional generalization Background knowledge Methodology 6.8.5. Zero-shot transfer learning for query suggestion Background knowledge Methodology 6.9. Summary References Chapter 7: Meta-reinforcement learning 7.1. Background knowledge 7.1.1. Basic components of a deep reinforcement learning system 7.1.2. Model-based and model-free approaches 7.1.3. Simulated environments 7.1.4. Limitations of deep reinforcement learning 7.2. Meta-reinforcement learning introduction 7.2.1. Early development 7.2.2. Formalism 7.2.3. Fundamental components 7.3. Memory 7.3.1. External read-write memory for agents with multiple modalities 7.4. Meta-reinforcement learning methods 7.4.1. Continuous adaptation in nonstationary environments Related Meta-RL algorithms for sample efficiency 7.4.2. Exploration with structured noise Related Meta-RL approaches for exploration 7.4.3. Credit assignment 7.4.4. Second-order computation in MAML Related Meta-RL algorithms based on MAML modifications 7.5. Reward signals and environments 7.5.1. Sparse extrinsic reward in procedurally generated environments Related Meta-RL algorithms for reward signal 7.6. Benchmark 7.6.1. Meta-World 7.7. Visual navigation 7.7.1. Introduction 7.7.2. Visual navigation to unseen scenes 7.7.3. Transferable meta-knowledge in unsupervised visual navigation 7.8. Summary References Chapter 8: Meta-learning for healthcare 8.1. Introduction Part I: Medical imaging computing 8.2. Image classification 8.2.1. Breast magnetic resonance imaging 8.2.2. Tongue identification 8.3. Lesion classification 8.3.1. Fine-grained skin disease classification 8.3.2. Difficulty-aware rare disease classification 8.3.3. Rare disease diagnostics: Skin lesion 8.4. Image segmentation 8.4.1. Medical ultra-resolution image segmentation 8.5. Image reconstruction 8.5.1. Chest and abdomen computed tomography image reconstruction Part II: Electronic health records analysis 8.6. Electronic health records 8.6.1. Disease prediction in a low-resource setting 8.6.2. Disease classification in a few-shot setting Part III: Application areas 8.7. Cardiology 8.7.1. Remote heart rate measurement in a few-shot setting 8.7.2. Customized pulmonary valve conduit reconstruction 8.7.3. Cardiac arrhythmia auto-screening 8.8. Disease diagnostics 8.8.1. Fine-grained disease classification under task heterogeneity 8.8.2. Clinical prognosis with Bayesian optimization 8.9. Data modality 8.9.1. Modality detection of biomedical images 8.10. Future work References Chapter 9: Meta-learning for emerging applications: Finance, building materials, graph neural networks, program synthesis ... 9.1. Introduction 9.2. Finance and economics 9.2.1. Introduction Approaches 9.2.2. Detection of credit card transaction fraud 9.2.3. Task-agnostic meta-learner with inequality measurement in economics Economic inequality measure 9.3. Building materials 9.3.1. Defect (crack) recognition in concrete in reinforcement learning 9.4. Graph neural network 9.4.1. Introduction 9.4.2. Node classification on graphs with few-shot novel labels 9.4.3. Local subgraphs for node classification and link prediction 9.4.4. Adversarial attacks of node classification Comparion and contrast of AQ and prototypical meta-learning 9.4.5. Dual-graph structured approach with instance- and distribution-level relations 9.5. Program synthesis 9.5.1. Syntax-guided synthesis 9.6. Transportation 9.6.1. Introduction 9.6.2. Traffic signal control 9.6.3. Continuous trajectory estimation for lane changes under a few-shot setting 9.6.4. Urban traffic prediction based on spatio-temporal correlation 9.7. Cold-start problems in recommendation systems 9.7.1. Introduction 9.7.2. Continuously adding new items 9.7.3. Context-aware cross-domain recommendation cold-start under a few-shot setting 9.7.4. User preference estimator 9.7.5. Memory-augmented recommendation system meta-optimization 9.7.6. Meta-learner with heterogeneous information networks 9.8. Climate science 9.8.1. Introduction 9.8.2. Critical incident detection 9.9. Summary References Index Back Cover