Deep Order-Preserving Learning With Adaptive Optimal Transport Distance.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Standard loss functions struggle to capture inherent label relationships and data correlations.
• Ordinal information, representing relative importance of labels, is often overlooked in predictive modeling.

Purpose of the Study:

• To develop a novel framework for learning label predictor functions that effectively incorporates label ordinality.
• To propose a new loss function for deep neural networks that leverages ordinal side information.

Main Methods:

• Utilized an optimal transport formulation to integrate label-data correlations.
• Learned a ground metric by incorporating ordinality as side information.
• Developed an efficient alternating learning algorithm for end-to-end optimization of the ground metric and deep model.

Main Results:

• The proposed method adaptively adjusts the loss function's shape based on application-specific needs.
• Demonstrated superior performance on chronological age estimation from faces and image aesthetic assessment tasks.
• Validated through theoretical analysis and numerical results on benchmark datasets.

Conclusions:

• The proposed optimal transport-based framework effectively leverages label ordinality for improved predictive modeling.
• The novel loss function and alternating learning scheme offer a flexible and powerful approach for deep learning applications.
• This method shows significant potential for enhancing various machine learning tasks requiring nuanced label understanding.