Goal driven network pruning for object recognition
作者：
Highlights：
• Proposed algorithm uncovers a smaller network competent in a sub-task in the dataset.
• It is inspired by the top-down attention part, removes non-contributory neurons.
• Eliminates re-training from scratch when a sub-task is assigned to the network.
• Hence, useful for the future’s unified networks trained on very multi-class datasets.
• Capable of reducing storage significantly and accelerating forward pass.
摘要
•Proposed algorithm uncovers a smaller network competent in a sub-task in the dataset.•It is inspired by the top-down attention part, removes non-contributory neurons.•Eliminates re-training from scratch when a sub-task is assigned to the network.•Hence, useful for the future’s unified networks trained on very multi-class datasets.•Capable of reducing storage significantly and accelerating forward pass.
论文关键词：Deep learning,Computer vision,Network pruning,Network compressing,Top-down attention,Perceptual visioning
论文评审过程：Received 14 July 2019, Revised 12 February 2020, Accepted 21 May 2020, Available online 27 May 2020, Version of Record 1 November 2020.
论文官网地址：https://doi.org/10.1016/j.patcog.2020.107468
