Co-Training Model Combination Pattern

A Co-Training Model Combination Pattern is a model combination pattern that uses two or more peer models to iteratively generate pseudo-labels for unlabeled data through different views or initializations.

• AKA: Co-Learning Pattern, Peer Teaching Pattern, Multi-View Learning Pattern.
• Context:
  • It can typically leverage Co-Training Unlabeled Data through co-training pseudo-label generations.
  • It can typically exploit Co-Training View Diversity via co-training complementary perspectives.
  • It can often improve Co-Training Model Agreement through co-training iterative refinements.
  • It can often reduce Co-Training Annotation Cost by co-training self-supervisions.
  • It can range from being a Two-Model Co-Training Pattern to being a Multi-Model Co-Training Pattern, depending on its co-training peer count.
  • It can range from being a View-Based Co-Training Pattern to being a Disagreement-Based Co-Training Pattern, depending on its co-training selection strategy.
  • It can range from being a Conservative Co-Training Pattern to being a Aggressive Co-Training Pattern, depending on its co-training confidence threshold.
  • It can range from being a Synchronous Co-Training Pattern to being a Asynchronous Co-Training Pattern, depending on its co-training update schedule.
  • ...
• Examples:
  • Classical Co-Training Patterns, such as:
    • Web Page Co-Training Pattern using page content view and hyperlink view.
    • Email Co-Training Pattern using header view and body view.
    • Document Co-Training Pattern using text view and metadata view.
  • Vision Co-Training Patterns, such as:
    • Multi-View Image Co-Training Pattern for 3D reconstruction.
    • RGB-Depth Co-Training Pattern for scene understanding.
    • Appearance-Motion Co-Training Pattern for video analysis.
  • NLP Co-Training Patterns, such as:
    • Bilingual Co-Training Pattern for machine translation.
    • Syntax-Semantics Co-Training Pattern for parsing.
    • Named Entity Co-Training Pattern for information extraction.
  • Multi-Modal Co-Training Patterns, such as:
    • Image-Text Co-Training Pattern for cross-modal learning.
    • Audio-Visual Co-Training Pattern for multimedia understanding.
  • ...
• Counter-Examples:
  • Self-Training Pattern, which uses a single model for pseudo-labeling.
  • Supervised Learning Pattern, which requires fully labeled data.
  • Active Learning Pattern, which queries human annotators rather than peer models.
• See: Model Combination Pattern, Semi-Supervised Learning, Self-Training Algorithm, Pseudo-Labeling Process, Multi-View Learning, Disagreement-Based Learning, Unlabeled Data Utilization.

---