Inaccurate-Supervised Learning With Generative Adversarial Nets.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Inaccurate-supervised learning (ISL) encompasses frameworks like partial label learning (PLL), partial multilabel learning (PML), and multiview PML (MVPML).
• Existing methods for ISL often treat these frameworks as independent problems and rely on traditional machine learning techniques.
• A unified framework, particularly one leveraging adversarial networks, was lacking for solving these diverse ISL problems.

Purpose of the Study:

• To propose a general adversarial network framework, termed ISL with generative adversarial nets (ISL-GANs), capable of addressing various ISL problems.
• To demonstrate the effectiveness of ISL-GANs in disambiguating noisy labels within imprecise annotation learning tasks.
• To provide theoretical analysis supporting the efficacy of ISL-GANs in handling ISL data.

Main Methods:

• Development of a novel adversarial network architecture (ISL-GANs) designed for inaccurate-supervised learning.
• Utilizing generative adversarial networks where generated fake samples aid in the discrimination of noisy labels in real samples.
• Formulating a general framework applicable to PLL, PML, and MVPML, with a specific conference version addressing PLL.

Main Results:

• ISL-GANs demonstrate effectiveness in solving PLL, PML, and MVPML tasks.
• The proposed framework successfully promotes label disambiguation in imprecise annotation scenarios.
• Extensive experiments validate the performance of ISL-GANs across various ISL tasks.

Conclusions:

• ISL-GANs offer a unified and effective adversarial framework for addressing multiple inaccurate-supervised learning problems.
• The method shows significant promise in improving the handling of imprecisely annotated data.
• This work bridges the gap by introducing the first general adversarial network framework for ISL problems.