[2018.05.11] Ensemble Soft-Margin Softmax Loss for Image Classification

[unizard]

IJCAI 2018 BK+ CS 우수국제학회 IF4.0되는 논문임

Institute: CBSR&NLPR URL: https://arxiv.org/pdf/1805.03922.pdf Author: http://www.cbsr.ia.ac.cn/users/xiaobowang/#publication, https://sites.google.com/view/xjguo/homepage Keyword: softmax discussion Interest: 1

결과만보면EM-SoftMax짱인데;; #NewClassificationLoss #Good

Backbone모델엄청가벼움 #진심후덜덜인데;;

1저자의기존연구결과가우수함 #본논문의신뢰성은보장됨

Summary

• This paper casts a new viewpoint on the weakness of softmax loss. On the one hand, the CNN features learned using the softmax loss are often inadequately discriminative.
• We hence introduce a soft-margin softmax function to explicitly encourage the discrimination between different classes.

[unizard]

Results

This method gives an more influence to large-scale DB.

1	2
	> ah.. Inference-Time consumes longer.
	> Wow... too light!!