Fast Broad Multiview Multi-Instance Multilabel Learning (FBM3L) With Viewwise Intercorrelation

Area of Science:

• Machine Learning
• Computer Vision
• Data Mining

Background:

• Multiview Multi-Instance Multilabel learning (M3L) is crucial for complex data like medical images and videos.
• Existing M3L methods face challenges with accuracy and training efficiency due to neglected correlations and high computational load.

Purpose of the Study:

• To propose a novel framework, fast broad M3L (FBM3L), addressing limitations of current M3L approaches.
• To enhance accuracy and training efficiency in M3L tasks, particularly for large-scale datasets.

Main Methods:

• Developed FBM3L framework utilizing viewwise intercorrelation, which was previously overlooked.
• Designed a viewwise subnetwork using Graph Convolutional Network (GCN) and Broad Learning System (BLS) for joint correlation learning.
• Leveraged the BLS platform for efficient joint learning across multiple views and subnetworks.

Main Results:

• FBM3L demonstrated highly competitive performance across all evaluation metrics, achieving up to 64% improvement in Average Precision (AP).
• The framework exhibited significant speed improvements, being up to 1030 times faster than existing M3L methods.
• FBM3L proved particularly effective on large multiview datasets, processing over 260,000 objects.

Conclusions:

• FBM3L offers a superior approach to M3L by effectively incorporating viewwise intercorrelations and diverse correlations.
• The proposed method significantly advances M3L by providing both high accuracy and exceptional training efficiency.
• FBM3L represents a substantial improvement for modeling complex real-world objects using multiview data.