KD-INR: Time-Varying Volumetric Data Compression via Knowledge Distillation-Based Implicit Neural Representation.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Traditional deep learning struggles with large-scale, time-varying data due to the need for all data during training.
• Effective data compression is crucial for managing and analyzing vast temporal datasets.

Purpose of the Study:

• To develop a novel data reduction pipeline for compressing large-scale time-varying data.
• To improve the efficiency of deep learning models handling dynamic datasets.

Main Methods:

• Proposed a two-stage pipeline: spatial compression using implicit neural representation (INR) and model aggregation via offline knowledge distillation.
• INR stage employs bottleneck layers and feature-preservation sampling for efficient time-step compression.
• Knowledge distillation aggregates knowledge from multiple trained models into a single, compressed model.

Main Results:

• KD-INR demonstrated superior performance over state-of-the-art learning-based and lossy compression methods.
• Achieved significant compression ratios, ranging from hundreds to ten thousand.
• Quantitative (PSNR, LPIPS) and qualitative (rendered images) evaluations confirmed KD-INR's effectiveness.

Conclusions:

• KD-INR offers a powerful solution for compressing large-scale time-varying data.
• The approach effectively handles the challenges posed by dynamic datasets in deep learning.
• KD-INR represents a significant advancement in data compression for complex, time-dependent information.