MoE-INR: Implicit Neural Representation with Mixture-of-Experts for Time-Varying Volumetric Data Compression.

Area of Science:

• Computer Vision
• Machine Learning
• Data Compression

Background:

• Implicit Neural Representations (INRs) are effective for high-dimensional signal modeling.
• Existing INR methods struggle with complex patterns and boundary artifacts.

Purpose of the Study:

• To introduce MoE-INR, an INR architecture using a mixture-of-experts framework.
• To address limitations of current INRs in modeling complex time-varying volumetric data.

Main Methods:

• Developed MoE-INR with a policy network, shared encoder, and expert decoders.
• The policy network automates field subdivision and expert assignment.
• Unified framework accommodates diverse INR types (conventional, grid-based, ensemble).

Main Results:

• MoE-INR significantly outperforms non-MoE and MoE-based INRs.
• Demonstrated superior performance over traditional lossy compression methods.
• Achieved better quantitative and qualitative metrics across various compression ratios.

Conclusions:

• MoE-INR offers a robust solution for time-varying volumetric data representation and compression.
• The mixture-of-experts approach enhances the modeling of complex spatiotemporal fields.
• MoE-INR represents a significant advancement in implicit neural representation techniques.