Computational-overhead characterization for compression and deployment

Characterize the computational overhead of dynamic-system-based neural-network model compression during both the compression process and model deployment.

Background

The proposed framework uses nearest-state assignment, KD-trees, coordinate templates, trajectory generation, and decompression from stored indices. Although these engineering techniques are intended to improve scalability, the paper does not provide a complete characterization of their computational overhead across compression and deployment. The authors explicitly list this characterization as an open question for future work.