Inductive State-Relabeling Adversarial Active Learning With Heuristic Clique Rescaling.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Active learning (AL) aims to reduce annotation costs by selecting the most representative samples.
• Existing AL methods struggle with insufficient annotations, unreliable uncertainty estimation, and sampling redundancy due to ignored intra-diversity.

Purpose of the Study:

• To propose an inductive state-relabeling adversarial AL model (ISRA) that enhances label efficiency and sample selection.
• To address the limitations of current AL approaches, including annotation inadequacy and sampling redundancy.

Main Methods:

• Developed an inductive state-relabeling adversarial AL model (ISRA) with a unified representation generator, inductive state-relabeling discriminator, and heuristic clique rescaling module.
• Incorporated contrastive learning for self-supervised training leveraging unlabeled data and mutual information to enhance representation quality.
• Designed an inductive uncertainty indicator for state scoring and relabeling unlabeled data, and a heuristic clique rescaling module to measure and manage intra-diversity of candidate samples.

Main Results:

• ISRA demonstrated superior performance compared to state-of-the-art AL methods across eight datasets and two imbalanced scenarios.
• The model effectively addresses sampling redundancy by measuring and rescaling intra-diversity of candidate samples.
• Achieved superior performance when applied to the cross-modal AL task of image captioning.

Conclusions:

• The proposed ISRA model offers a significant advancement in label-efficient algorithms for active learning.
• ISRA effectively overcomes key challenges in AL, leading to improved sample selection and reduced annotation costs.
• The model's adaptability extends to cross-modal tasks, showcasing its versatility and effectiveness.