Structured deep Fisher pruning for efficient facial trait classification
作者：
Highlights：
• Large networks are usually over-parameterized.
• Optimal deep net structures are task-specific.
• Networks can be effectively pruned via linear discriminant analysis in the feature (map) space.
• Last conv layer filters in a fine-tuned facial trait recognition CNN tend to fire uncorrelatedly.
• Pruning can possibly boost network accuracy in addition to bringing down complexity.
摘要
•Large networks are usually over-parameterized.•Optimal deep net structures are task-specific.•Networks can be effectively pruned via linear discriminant analysis in the feature (map) space.•Last conv layer filters in a fine-tuned facial trait recognition CNN tend to fire uncorrelatedly.•Pruning can possibly boost network accuracy in addition to bringing down complexity.
论文关键词：Neural network pruning,Fisher LDA,Facial trait classification
论文评审过程：Received 16 September 2017, Revised 4 June 2018, Accepted 21 June 2018, Available online 5 July 2018, Version of Record 20 July 2018.
论文官网地址：https://doi.org/10.1016/j.imavis.2018.06.008
