Imbalanced Deep Learning by Minority Class Incremental Rectification.

Area of Science:

• Machine Learning
• Deep Learning
• Computer Vision

Background:

• Class imbalanced data presents a significant challenge in machine learning model training.
• Existing deep learning methods often fail to adequately address significantly imbalanced datasets, focusing instead on balanced or moderately imbalanced data.

Purpose of the Study:

• To develop a deep learning model capable of effectively learning from significantly imbalanced training data.
• To minimize the dominant influence of majority classes in imbalanced datasets.

Main Methods:

• Formulation of a class imbalanced deep learning model utilizing batch-wise incremental minority class rectification.
• Implementation of hard sample mining within majority classes during iterative batch-wise learning.
• Introduction of a novel Class Rectification Loss (CRL) function compatible with deep network architectures.

Main Results:

• The proposed model demonstrates performance advantages over state-of-the-art methods on imbalanced person attribute datasets (CelebA, X-Domain, DeepFashion).
• Experimental results confirm the model's scalability and effectiveness in addressing imbalanced data learning challenges.
• The Class Rectification Loss (CRL) function proved readily deployable and effective.

Conclusions:

• The batch-wise incremental minority class rectification model effectively addresses the challenge of learning from significantly imbalanced data.
• The proposed CRL function and model architecture offer a scalable solution for imbalanced deep learning.
• This approach enhances model performance by discovering minority class boundaries through iterative learning.