Multimodel Feature Reinforcement Framework Using Moore-Penrose Inverse for Big Data Analysis.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fully connected representation learning (FCRL) is common in multimodal image classification.
• Existing FCRL methods often use separate blocks, leading to loosely connected feature representations.

Purpose of the Study:

• To propose a novel hierarchical subnetwork-based neural network (HSNN) for robust feature representation.
• To achieve a more robust representation by simultaneously considering low-dimensional features and the classifier model.

Main Methods:

• An iterative learning process that integrates feature encoding and classification.
• A Moore-Penrose (MP) inverse-based batch-by-batch learning strategy for handling large datasets.
• The proposed HSNN framework.

Main Results:

• The HSNN framework generates optimal global features through iterative learning.
• Effective processing of large-scale datasets, such as Place365 (1.8 million images).
• Demonstrated superior generalization performance across various dataset sizes compared to state-of-the-art FCRL methods.

Conclusions:

• The proposed HSNN framework offers a more effective approach to representation learning.
• Iterative learning and MP inverse-based strategies enhance performance and scalability.
• The HSNN framework achieves better generalization in image classification tasks.