Inhalt

•  
   1 Introduction
•  
   1.1 Characteristics of life-long learning
•  
   1.2 Technical challenges
•  
   1.3 A basic model for life-long learning
•  
   1.4 Contributions
•  
   1.5 Components of GCM (Overview)
•  
   1.6 Outline
•  
   2 Related Work
•  
  2.1 Topological feature maps
•  
   2.1.1 Growing Neural Gas (GNG)
•  
   2.1.2 Extensions of GNG
•  
  2.2 Learning with continuous data streams
•  
   2.2.1 Challenges & Methodologies
•  
   2.2.2 Domain-specific solutions
•  
  2.3 Learning with weak supervision
•  
   2.3.1 Categories of SSL Methods
•  
   2.3.2 Data assumptions
•  
   2.3.3 Weak supervision and late labeling
•  
  2.4 Learning with non-stationary data
•  
   2.4.1 Challenges & Methodologies
•  
   2.4.2 Layering of memory
•  
   2.5 Learning with multi-view data
•  
   2.6 Summary
•  
   3 Classification of continuous data streams based on OGNG
•  
   3.1 Challenges
•  
   3.2 Contributions
•  
   3.3 Classification with GNG
•  
   3.3.1 Offline labeling strategies
•  
   3.3.2 Prediction strategies for GNG
•  
   3.3.3 Limitations of batch learning
•  
  3.4 Online Growing Neural Gas (OGNG)
•  
   3.4.1 Online labeling strategies for GNG
•  
   3.4.2 OGNG Algorithm
•  
   3.4.3 A uniform two-level architecture
•  
   3.5 Datasets
•  
  3.6 Evaluation
•  
   3.6.1 Parameters & OGNGtop
•  
   3.6.2 Experiments & Results
•  
   3.6.3 Discussion
•  
   3.7 Summary
•  
   4 Learning with weak supervision based on OSSGNG
•  
   4.1 Challenges
•  
   4.2 Contributions
•  
   4.3 Offline semi-supervised learning with GNG
•  
   4.3.1 Semi-supervised Growing Neural Gas (SSGNG)
•  
   4.3.2 SSGNG Algorithm
•  
   4.3.3 Limitations of SSGNG
•  
  4.4 Online Semi-supervised Growing Neural Gas (OSSGNG)
•  
   4.4.1 OSSGNG as extension of OGNG
•  
   4.4.2 OSSGNG Algorithm
•  
   4.4.3 Dynamics of the OSSGNG labeling
•  
   4.5 Datasets
•  
   4.6 Evaluation
•  
   4.6.1 Parameters
•  
   4.6.2 Experiments & Results
•  
   4.6.3 Discussion
•  
   4.7 Summary
•  
   5 Learning with non-stationary data based on DYNG
•  
   5.1 Challenges
•  
   5.2 Contributions
•  
   5.2.1 Growing Neural Gas with Utility (GNG-U)
•  
   5.2.2 GNG-U Algorithm
•  
   5.2.3 Limitations of GNG-U
•  
  5.3 Dynamic Online Growing Neural Gas (DYNG)
•  
   5.3.1 DYNG as extension of OGNG
•  
   5.3.2 DYNG Algorithm
•  
   5.3.3 Memory Structures
•  
   5.4 Dataset
•  
  5.5 Evaluation
•  
   5.5.1 Parameters & OGNG-U
•  
   5.5.2 Experiments & Results
•  
   5.5.3 Discussion
•  
   5.6 Summary
•  
   6 Learning with multi-view data based on VONG
•  
   6.1 Challenges
•  
   6.2 Contributions
•  
  6.3 OGNG+: A naive multi-label approach based on OGNG
•  
   6.3.1 OGNG+ Algorithm
•  
   6.3.2 Limitations of OGNG+
•  
   6.4 Multi-view Online Growing Neural Gas (VONG)
•  
   6.4.1 VONG as multi-view extension of OGNG
•  
   6.4.2 VONG Algorithm
•  
   6.4.3 Multiplexing of views
•  
   6.5 Datasets
•  
  6.6 Evaluation
•  
   6.6.1 Parameters & VONG-M
•  
   6.6.2 Experiments & Results
•  
   6.6.3 Discussion
•  
   6.7 Summary
•  
   7 Applications for Growing Conceptual Maps
•  
  7.1 Classification of textual data in social media streams
•  
   7.1.1 The TwitterLL corpus
•  
   7.1.2 The Reuters RCV1 v2 corpus
•  
   7.1.3 Experiments & Parameters
•  
   7.1.4 Results & Discussion
•  
   7.2 Human activity classification
•  
   7.2.1 Human Activity Classification with OGNG
•  
   7.2.2 KTH human motion dataset
•  
   7.2.3 Experiments & Results
•  
   7.2.4 Discussion
•  
  7.3 Visualization of Conceptual Maps
•  
   7.3.1 A visualization schema for Growing Conceptual Maps
•  
   7.3.2 A generic visualizations concept
•  
   7.3.3 Examples of visualized Conceptual Maps
•  
   7.3.4 Discussion
•  
   7.4 Summary
•  
   8 Conclusion
•  
   A Conceptual Map Visualizations
•  
   B Bibliography