Deep learning from label proportions with labeled samples.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Learning from Label Proportions (LLP) uses bag-level class proportion data.
• Existing LLP methods often ignore valuable labeled samples within bags.
• This oversight limits the potential for capturing essential data distribution information.

Purpose of the Study:

• To propose an end-to-end LLP solver, LLP with Labeled Samples (LLP-LS), that effectively utilizes labeled samples.
• To enhance the performance of LLP algorithms by integrating proportional and labeled data.
• To achieve state-of-the-art results in multi-class classification problems.

Main Methods:

• Developed a novel LLP solver, LLP-LS, based on Convolutional Neural Networks (ConvNets).
• Reshaped the cross-entropy loss function to incorporate both proportional and labeled sample information from each bag.
• Employed batch-based ADAM optimization, aligning batch size with bag size for effective training.

Main Results:

• The proposed LLP-LS method demonstrated superior performance compared to existing state-of-the-art methods.
• Achieved top-tier results on several benchmark image datasets for multi-class classification.
• Effectively leveraged labeled samples within bags, improving data distribution information utilization.

Conclusions:

• LLP-LS offers a significant advancement in Learning from Label Proportions.
• The method successfully integrates proportional and labeled data for enhanced model training.
• LLP-LS sets a new benchmark for performance in LLP tasks, particularly for image datasets.