HAda: Hyper-Adaptive Parameter-efficient Learning for Multi-view ConvNets.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep Convolutional Neural Networks (ConvNets) have shown success in multi-view learning but often require a large number of parameters.
• Hypernetworks offer a method to reduce parameter count by generating weights for target networks, indicating parameter redundancy in existing models.

Purpose of the Study:

• To address the underexplored area of leveraging hypernetworks for parameter-efficient multi-view ConvNets.
• To develop a lightweight network that generates adaptive weights for different views and convolutional layers, reducing redundancy and maintaining performance.

Main Methods:

• Proposed a lightweight multi-layer shared Hyper-Adaptive network (HAda).
• Designed a multi-view shared module with a gated interpolation strategy for adaptive weight generation.
• Introduced view-specific weight-calibrated adapters for personalized information emphasis.

Main Results:

• HAda achieved substantial reduction in parameter redundancy.
• The method effectively modeled intricate view-aware and layer-wise information, maintaining high performance.
• Extensive experiments on six datasets demonstrated the effectiveness of HAda for image classification and clustering.

Conclusions:

• HAda offers a parameter-efficient solution for deep multi-view ConvNets.
• The proposed hypernetwork approach successfully balances parameter reduction with performance maintenance.
• HAda can be flexibly integrated as a plug-in strategy with existing multi-view methods.