There are quite a lot of God-made papers in this issue, and they are very interesting.
Feature representation in Convolutionalneural Networks
In this paper, it is discussed in a certain kind of CNN structure, whether there is a higher accuracy of the off Model classification method (here refers to non-SOFTMAX) can achieve more effective classification results?
The paper gives a definite answer.
This paper also gives a chart of the importance of each layer, quite interesting
The paper also gave an account of the open source code used in the experiment.
Towards good practices for Very Deeptwo-stream Convnets
Openmpi for Multi-gpu
Code:https://github.com/yjxiong/caffe/tree/action_recog
Two-stream convolutional Networks foraction recognition in Videos
The accuracy of the single stream is as follows:
The results of the combine are as follows:
Understanding Intra-class Knowledge INSIDECNN
This paper uses visualization techniques to describe how CNN is differentiated within a class.
In addition, there are references to the use of visual representations of how CNN can differentiate between classes.
is a good article on CNN visual comprehension
Compressing deep convolutional networksusing VECTOR quantization
This paper describes the mobile-level model storage compression, a very good paper.
In fact, the compression of model parameters will not only play a role in storage space compression (PQ), but also play a role in accelerating the model such as SVD.
Unconstrained facial Landmark localizationwith backbone-branches fully-convolutional Networks
The paper puts forward the network structure of backbone-branches fully-convolutional Neural Networks (BB-FCN), which is very interesting, and gives a lot of commercial non-commercial methods to compare, is a very good face positioning cut into the article.
The network structure is as follows:
Results Comparison table:
Facial Landmark Detection by Deepmulti-task Learning
Multi-tasking learning can improve the accuracy of a single task.
Copyright NOTICE: This article for Bo Master original article, without Bo Master permission not reproduced.
[Thu, 9 Jul ~ Tue, 2015 Jul] Deep Learning in arxiv