Learning by Seeing More Classes.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional pattern recognition models assume a fixed number of classes during training and inference.
• The impact of increasing class numbers during training on model performance is largely unexplored.

Purpose of the Study:

• To investigate whether training with additional classes improves generalization and reliability in pattern recognition.
• To introduce and evaluate a novel approach called class augmentation (classAug).

Main Methods:

• Proposed learning with k+m classes instead of only k classes for a k-class problem.
• Developed two strategies for synthesizing new classes from existing ones.
• Compared class augmentation (classAug) with traditional data augmentation (dataAug).

Main Results:

• Training with additional classes acts as regularization, improving generalization accuracy on original classes.
• Reduced model overconfidence, leading to more reliable confidence estimation for tasks like misclassification and out-of-distribution detection.
• Enhanced feature representation beneficial for few-shot and class-incremental learning.

Conclusions:

• Class augmentation (classAug) offers a new dimension for improving model performance beyond data augmentation.
• The proposed classAug method demonstrates superiority across various open-environment metrics on benchmark datasets.