講演抄録/キーワード |
講演名 |
2022-09-15 10:00
One-cut Network Pruning at Initialization with Explainable Image Concepts ○Yinan Yang(Rits Univ.)・Ying Ji(Nagoya Univ.)・Yu Wang(Hitotsubashi Univ.)・Jien Kato(Rits Univ.) PRMU2022-18 |
抄録 |
(和) |
(まだ登録されていません) |
(英) |
Recent research investigates the feasibility of one-cut network pruning at initialization (OPAI). SNIP and GraSP are two representative OPAI methods. Recent work asserts SNIP and GraSP are data-independent as they are robust with random corrupting labels or pixels. It indicates that OPAI ``hardly exploits any information from training data''. However, we obtain an opposite conclusion by proposing Concept-based One-cut Network Pruning (COP) and Super Stitching sampling strategy. COP and Super Stitching extract high-level visual concepts from the dataset with ranking and then prunes the network by these concepts. We find that: (1) The COP and ablation experiments reveal that the OPAI is data dependent. (2) Super Stitching performs better than the original OPAI method on benchmark ImageNet using concepts rather than random images. COP and Super Stitching, in contrast to existing human-unfriendly pruning methods, employ genuine explainable visual concepts rather than blind pruning. |
キーワード |
(和) |
/ / / / / / / |
(英) |
Network Pruning / Discriminative / Explainable Image Concepts / / / / / |
文献情報 |
信学技報, vol. 122, no. 181, PRMU2022-18, pp. 49-54, 2022年9月. |
資料番号 |
PRMU2022-18 |
発行日 |
2022-09-07 (PRMU) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
PRMU2022-18 |