Incremental Learning for Simultaneous Augmentation of Feature and Class.

Area of Science:

• Machine Learning
• Computer Vision
• Pattern Recognition

Background:

• Dynamic environments generate data with gradually accumulating features and increasing class numbers.
• Existing methods struggle with simultaneous feature and class augmentation, especially with limited labeled data.
• Activity recognition exemplifies scenarios requiring adaptation to new sensors and exercise types.

Purpose of the Study:

• To propose a novel incremental learning method for Simultaneous Augmentation of Feature and Class (SAFC).
• To address the challenge of learning with simultaneously augmenting features and classes in dynamic environments.
• To ensure model reusability and validate theoretical efficiency for evolving datasets.

Main Methods:

• A two-stage incremental learning approach for SAFC is proposed.
• A regularizer is incorporated to ensure reusability of models trained on previous data.
• Theoretical analyses of generalization bounds are presented to validate model inheritance efficiency.
• The method is extended from one-shot to multi-shot learning scenarios.

Main Results:

• The proposed SAFC method demonstrates effectiveness in handling simultaneous feature and class augmentation.
• The regularizer aids in training new classifiers by providing a solid prior from previous data.
• Theoretical analyses confirm the efficiency of model inheritance.
• Experimental results validate the approach's performance, particularly in activity recognition.

Conclusions:

• The SAFC method offers a robust solution for incremental learning in dynamic environments with evolving data and classes.
• The approach is effective even with limited labeled samples and sensor data.
• The method shows significant promise for real-world applications like activity recognition.