Semantic-Aware Lossless Floating-Point Compression With Parallel Recovery for Intelligent Systems
International Journal on Semantic Web and Information Systems · 12.11.2026 · 22 (1) · 1-32
DOI: 10.4018/ijswis.427371 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal on Semantic Web and Information Systems, erschienen am 12.11.2026. Verfasst von Changyu Gong und 1 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 3.0).
Abstract (Originalsprache)
Detectors in intelligent information systems are often constrained by model storage, graphics processing unit (GPU) memory, and weight-transfer bandwidth. Pruning, quantization, and weight-sharing methods usually alter the network structure or weight values and require retraining or calibration. Existing lossless compressors are generally designed for model-file storage and still require complete decompression before inference. The authors propose LEAD-Float, a semantic-aware lossless compression method that uses lightweight semantic metadata to organize local exponent coding and block-addressable recovery. Structure-aware local Huffman codebooks capture the heterogeneous exponent distributions of component weights, while independently addressable blocks support parallel recovery at convolutional access granularity. A selective recovery–computation fusion path further avoids full-model decompression and unnecessary intermediate GPU-memory writeback. Across 14 detectors, it reduces fusible-weight storage to 68.12% for BF16 and 86.67% for FP16 and improves recovery throughput by 34.43% and 52.37% over the baseline.
Zitieren
APA
Gong, C., & Zhao, Y. (2026). Semantic-Aware Lossless Floating-Point Compression With Parallel Recovery for Intelligent Systems. International Journal on Semantic Web and Information Systems, 22(1), 1-32. https://doi.org/10.4018/ijswis.427371
Vancouver
Gong C, Zhao Y. Semantic-Aware Lossless Floating-Point Compression With Parallel Recovery for Intelligent Systems. International Journal on Semantic Web and Information Systems. 2026;22(1):1-32. doi:10.4018/ijswis.427371
BibTeX
@article{Gong2026,
author = {Gong, Changyu and Zhao, Yafeng},
title = {Semantic-Aware Lossless Floating-Point Compression With Parallel Recovery for Intelligent Systems},
journal = {International Journal on Semantic Web and Information Systems},
year = {2026},
volume = {22},
number = {1},
pages = {1--32},
doi = {10.4018/ijswis.427371}
}
EUVIMED