Sample Balancing for Deep Learning-Based Visual Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Sample balancing, encompassing sample selection and reweighting, is crucial for optimizing deep learning model performance.
• Existing sample reweighting methods often overlook sample quality, relying primarily on per-class sample counts.

Purpose of the Study:

• To investigate the impact of diverse sample selection strategies on deep network training.
• To develop an improved sample reweighting method that considers sample quality.
• To integrate these methods into deep learning frameworks for end-to-end training.

Main Methods:

• Incorporation of a self-paced learning-based sample selection technique into deep learning frameworks.
• Proposal of a novel sample reweighting metric utilizing multiview semantic encoders.
• Development of an optimization mechanism to embed sample weights into deep network loss functions for end-to-end training.

Main Results:

• Experimental validation on CIFAR and ImageNet datasets.
• Demonstrated improvement in the performance of deep learning methods across various visual recognition tasks.
• The proposed sample balancing method effectively enhances deep learning model accuracy.

Conclusions:

• The integrated sample balancing approach, combining advanced sample selection and reweighting, significantly boosts deep learning performance.
• The novel multiview semantic encoder-based reweighting metric offers a more effective way to utilize sample information.
• This method provides a robust strategy for improving deep learning in visual recognition.