A Transfer Learning-Based Multi-Instance Learning Method With Weak Labels.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multi-instance learning (MIL) typically assumes bags have true labels, which is often not the case in real-world scenarios.
• Weak labels arise from aggregating multiple labeler inputs with unknown weights, leading to ambiguity in MIL.
• Knowledge transfer between related tasks (multiple instance transfer learning) is crucial but challenging with weak labels.

Purpose of the Study:

• To propose a novel Transfer Learning-based Multiple Instance Learning (TMIL) framework.
• To address the challenge of multiple instance transfer learning when both source and target tasks have weak labels.
• To improve the performance of MIL by effectively transferring knowledge and correcting ambiguous bag labels.

Main Methods:

• Developed a TMIL model designed to handle weak labels in both source and target tasks.
• Introduced an iterative framework to solve the transfer learning model with weak labels.
• Performed convergence analysis for the proposed iterative method.

Main Results:

• The proposed TMIL framework successfully transfers knowledge from source to target tasks with weak labels.
• The iterative approach refines initial weak bag labels, leading to improved MIL performance.
• Experimental results demonstrate that the TMIL method outperforms existing MIL approaches.

Conclusions:

• The TMIL framework provides an effective solution for multiple instance transfer learning with weakly labeled data.
• The iterative label updating mechanism enhances the accuracy of MIL by correcting ambiguous labels.
• The proposed method shows significant improvements over traditional MIL techniques.