|
|
© 2023 Meta-Learning in Computer Vision Edit: Prof. Dr. Wen-Feng Wang, Lalit Mohan Patnaik, Fuqing Li Financially supported by the Shanghai High-level Base-building Project for Industrial Technology Innovation (1021GN204005-A06). ISBN: 978-1-80430-002-2 DOI: 10.56157/ETSL5404 Publication Date: November 15, 2023 |
Description
Early machine learning was not end-to-end, that is, it did not directly input raw data at the input end, but based on human prior knowledge, preprocessed the raw data and used the extracted features as input. So early machine learning was also known as feature engineering. Due to the fact that the input raw data is manually processed before being used to train the model, it is impossible to avoid situations such as incorrect feature selection, inaccurate feature extraction, and even significant deviation in feature calculation results.
With the development of machine learning, especially the emergence of multi-layer neural networks, the end-to-end learning process has been achieved, marking the transition of machine learning from feature engineering to representation learning. Deep learning is further developed on this basis. Deep learning models have the ability to learn features. Because by configuring relevant parameters, multi-layer feedforward neural networks can approximate any function.
Metalearning belongs to the category of machine learning. The ultimate goal of machine learning algorithm research is to enable machines to have learning abilities that are close to those of humans. This is a very big challenge! The current mainstream machine learning methods, whether classical machine learning theories or more advanced deep learning models, cannot achieve this. The emergence of meta learning is precisely to enable machines to have learning abilities that are close to those of humans.
Machine learning theory is evolving from "mechanical memory" to "active learning and application", which is an inevitable trend. Just like a person's pursuit of a student career. At the beginning, we had relatively little consideration for learning methods, mainly relying on the matching of memory and knowledge to complete our preliminary understanding of the world. When the difficulty of knowledge increases, simple memory and matching can no longer meet the needs of learning, we have to start to pay attention to learning methods. This book will provide a learning-to-learn theory for the machine brain, which will help overcome the limitations of existing machine learning algorithms.
About the authors
Wen-Feng Wang is a full professor in Shanghai Institute of Technology. He is a key talent in Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences (2018- ), a reviewer of many SCI journals, including some top journals - Water Research, Science China-Information Sciences, Science of the Total Environment, Environmental Pollution, IEEE Transactions on Automation Science and Engineering and etc. He served as a keynote speaker of AMICR2019, IACICE2020, OAES2020, AICS2021, 3DIT-MSP&DL2020 and etc. He is also the scientist in chief of RealMax in the research field of intelligent computing and big data.
Lalit Mohan Patnaik has served as the President of the Advanced Computing and Communications Society, and the Computational Intelligence Society of India. He is a Fellow of The World Academy of Sciences, Indian National Science Academy, Indian Academy of Sciences, National Academy of Sciences, Indian National Academy of Engineering, Maharashtra Academy of Sciences. His name appears in Asia’s Who's Who of Men and Women of Achievement; Directory of Distinguished Computer Professionals of India. For over one hundred and seventy International Conferences, he has served as General Chair or Program Chair or Steering Committee Chair; most of them sponsored by the Institute of Electrical and Electronics Engineers (IEEE).
Fuqing Li is a professor and doctoral supervisor of Xiangtan University. She was nominated as a team leader of national woman civilization station and the scholar of the 121 Innovative Talent Projects and Lotus Scholar Award Program in Hunan Province. She served as a senior researcher at the University of Connecticut in USA and finished her post-doctoral program and advanced visiting scholar program in Xiangtan University and Philosophy Institute of Chinese Academy of Social Sciences respectively. She obtained multiple research awards and won numerous teaching competitions at provincial and ministerial level, constructed the first-class virtual simulation course and “the Golden Course" in Hunan Province. Prof. Fuqing Li hosted three national social science funds projects and eight provincial and ministerial projects. She published five monographs in top publishers and more than 60 papers in CSSCI indexed or other prominent journals. Her research interests include digital course construction and teaching reform, development and construction of Virtual Simulation resources for Humanities and the study of Modern Chinese thought.
Associate Editor in Chief: Fa-Ming Huang
Dr. Huang is an outstanding young scholar who is striving for excellence in risk assessment of geological hazards using “3S” technologies, machine learning models and geotechnical numerical simulation. Dr. Huang now serves as the Associate Editors of the SCI indexed journal of “Frontiers in Earth Science” and CSCD indexed journal of “Bulletin of Geological Science and Technology”. He is also invited as the editorial board member in the journals of “International Journal of Applied Nonlinear Science”, “Journal of Environmental & Earth Sciences” and “Natural Hazards Research”, as well as the Youth editorial board member in the journal of “International Journal of Mining Science and Technology”, “Rock Mechanism Bulletin” and “International Journal of Coal Science & Technology”. He also served as the Lead guest editor of the special issue “Remote Sensing for Geohazards Monitoring: Towards Refined Risk Analysis and Management” in the journal “Remote Sensing” in 2022. Furthermore, in 2022 and 2023 years, He served as the guest editor and main contact of the special issue “Machine learning and Engineering risk assessment” in the EI indexed journals of “Earth Science” and “Journal of Basic Science and Engineering”.
Table of Contents
Meta-Learning in Computer Vision
Editorial Committee
Preface
Chapter 1. The "Learn-to-Learn" Problem
1.1 Introduction
1.2 Problem Formulation
1.2.1 Formulation of meta-learning
1.2.2 Meta-training and meta-testing
1.2.3 Formulation of meta-learning tasks
1.3 The brain functional activities
1.3.1 The concept of biosignals
1.3.2 Functions for convolution calculation
1.4 Summary and Programming
Chapter 2. Prototypical Network Structure
2.1 Introduction
2.2 Problem Formulation
2.2.1 Principles for learning
2.2.2 Prototypical Representations
2.3 Implementation of the Network
2.3.1 Data set download and import
2.3.2 Realization of the training process
2.3.3 Resolution of the difficult issues
2.4 Summary and Programming
Chapter 3. Learning to Modify Features
3.1 Introduction
3.2 Problem Formulation
3.2.1 Modeling approaches
3.2.2 Mechanisms of modification
3.2.3 Self-directed characterization
3.3 Core Ideas of Autonomous Learning
3.3.1 Graph-structured data
3.3.2 Perturbation of images
3.3.3 Two-layer optimization problem
3.3.4 Future Directions for Attack Algorithms
Chapter 4. Interpretation of Meta-learners
4.1 Introduction
4.2 Problem Formulation
4.2.1 LSTM meta-learning system
4.2.2 MAML Algorithm
4.2.3 LSTM-based meta learner
4.3 Few-shot learning and meta learning
4.3.1 Few-shot learning principles
4.3.2 Metric-based meta learning algorithms
4.4 Meta-learning using LSTM
4.4.1 Basic steps of LSTM
4.4.2 Training a meta-learner
4.5 Convex Optimization Meta-Learning
4.5.1 Proposed Learning Framework
4.5.2 learning process
4.5.3 base learner selection
4.5.4 Acquisition of Embedded Models
4.5.5 Formulation of the problem
4.5.6 Optimization of the model
4.5.7 Program framework
Chapter 5. Reinforced Learning with Meta-idea
5.1 Introduction
5.2 Problem Formulation
5.2.1 Important tips for DQN
5.2.2 Bellman equation and state value function
5.3 Combining DQN with meta learning
5.3.1 Basic steps of entropy weight method
5.3.2 Meta LSTM optimizer
5.3.3 MQN network architecture
5.4 Summary and Conclusions
Chapter 6. Grap of Meta learning Frameworks
6.1 Introduction
6.2 Problem Formulation
6.2.1 Graph CNN
6.2.2 Properties and characteristics
6.2.3 The Entire Graph of GCN
6.3 A meta-learning framework based on GCN
6.3.1 The fundamental idea
6.3.2 Classification and Advancements
6.3.3 The Integration Approach of GCN
6.4 Summary and Conclusions
Chapter 7. Meta Learning in Optimization
7.1 Introduction
7.2 Problem Formulation
7.2.1 Definition of parameter optimization
7.2.2 Methods for parameter optimization
7.3 Parameter optimization in meta-learning
7.3.1 Meta-learning based on Memory
7.3.2 Meta-learning based on predicted gradients
7.3.3 Meta-learning using attention mechanisms
7.3.4 Borrow from the meta-learning of LSTM
7.3.5 Meta-learning for reinforcement learning
Chapter 8. Prototypical R einforcement Network
8.1 Introduction
8.2 Problem Formulation
8.2.1 Prototype representation
8.2.2 Algorithm description
8.2.3 Network Expansion
8.2.4 Model optimization
8.3 Drawing on enhanced learning
8.3.1 Learning styles
8.3.2 Network architecture
8.3.3 Model description
Chapter 9. Hypotheses for Joint Tasks Training
9.1 Introduction
9.2 Problem Formulation
9.2.1 The algorithm idea of meta learning
9.2.2 Mathematical expression of the basic model
9.3 Joint training optimization problems in 3-element learning
9.3.1 Basic hypothesis and analysis of the problem
9.3.2 Derivation and results comparison
9.3.3 Details in the optimization process
Chapter 10. Improved Meta learning Architecture
10.1 Introduction
10.2 Problem Formulation
10.2.1 Concept of meta-learning
10.2.2 Meta-learning architecture
10.2.3 The idea of meta-learning method
10.2.4 Training process of learning
10.3 Model independent meta-learning
10.3.1 Introduction to model independent meta-learning
10.3.2 Main algorithms of the meta optimizer
10.4 Use LSTM meta-learning
10.4.1 Recurrent neural network RNN
10.4.2 Long Short-Term Memory (LSTM) network
10.4.3 Work steps of LSTM in meta-learning
Chapter 11. Further Innovation of Meta learners
11.1 Introduction
11.2 Problem Formulation
11.2.1 The origin of meta-learning
11.2.2 Applications of meta-learning
11.3 Machine learning and meta-learning
11.3.1 The concept distinction
11.3.2 Machine learning and Meta-learning steps
11.3.3 Differences from Machine Learning
11.4.LSTM as a meta-learner learns gradient descent
11.4.1 LSTM model and LSTM under Meta-Learning
11.4.2 Learn more about LSTM
11.4.3 Parameter sharing and preprocessing
11.4.4 Gradient independent thinking
11.4.5 Model training
11.4.6 Innovation of LSTM as a meta-learner
Chapter 12. Applications in Computer Vision
12.1 Introduction
12.2 Problem Formulation
12.2.1 Basics of K-means clustering algorithm
12.2.2 Idea of K-means clustering algorithm
12.2.3 Flow of K-means clustering algorithm
12.2.4 Problem Analysis and Existing Improvements of K-means Clustering Algorithm
12.3 Combined application of K-means clustering algorithm and metric learning
12.3.1 Meta-learning and clustering algorithms
12.3.2 Image segmentation based on K-means clustering algorithm
References