Classification and comparison via neural networks.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Traditional machine learning often relies on class labels, which can be high-variance.
• Comparison labels, indicating relative order between samples, offer lower variance but scale quadratically.
• Integrating diverse label types can enhance model performance and data efficiency.

Purpose of the Study:

• To develop a neural network architecture capable of utilizing both class and comparison labels.
• To improve the accuracy of classification and comparison label prediction.
• To demonstrate the feasibility of training deep neural networks with limited data using comparison labels.

Main Methods:

• Proposed a novel neural network architecture based on siamese networks.
• Incorporated both class and comparison labels within a unified training pipeline.
• Utilized Bradley-Terry and Thurstone loss functions for model training.

Main Results:

• Achieved significant improvements in predicting both class and comparison labels.
• Increased classification AUC by up to 35% and comparison AUC by up to 6% on real-life datasets.
• Demonstrated successful training of a large neural network on minimal data (80 images) by incorporating comparisons, reaching 0.92 AUC.

Conclusions:

• The proposed siamese network architecture effectively integrates comparison labels with class labels.
• Comparison labels significantly enhance model performance and enable few-shot learning scenarios.
• This approach offers a promising direction for improving machine learning models in data-scarce environments.