A Novel Structure-Agnostic Multi-Objective Approach for Weight-Sharing Compression in Deep Neural Networks
Rasa Khosrowshahli, Student Member, IEEE, Shahryar Rahnamayan, Senior Member, IEEE, Beatrice Ombuki-Berman, Senior Member, IEEE · 2025
Read A Novel Structure-Agnostic Multi-Objective Approach for Weight-Sharing Compression in Deep Neural Networks by Rasa Khosrowshahli, Student Member, IEEE, Shahryar Rahnamayan, Senior Member, IEEE, Beatrice Ombuki-Berman, Senior Member, IEEE (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-03095.txt.
Description
Rasa Khosrowshahli, Shahryar Rahnamayan, and Beatrice Ombuki-Berman wrote this paper in two thousand twenty-five. It squeezes deep neural nets by sharing weights, with a search method that does not care what shape the net has.
Big nets store vast counts of weights, which strains small devices. The frame bins all weights into one shared codebook with even bin widths, then runs a multi-objective evolutionary search for the best trade of accuracy against bin count. A merge step then fuses near bins on the best front to shrink the book more. Cluster centers serve as the shared weights, so no costly retraining is needed. Tests cut memory by more than tenfold on standard image sets.
Catalog record
| CREATOR | Rasa Khosrowshahli, Student Member, IEEE, Shahryar Rahnamayan, Senior Member, IEEE, Beatrice Ombuki-Berman, Senior Member, IEEE |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-03095