[Fri, 11 Dec 2015~ Fri, 25 Dec 2015] Deep Learning in arxiv

EmotionRecognition in the Wild via Convolutional Neural Networks and Mapped BinaryPatterns

Caffemodel zoo里的一个分享。

Paper:http://www.openu.ac.il/home/hassner/projects/cnn_emotions/LeviHassnerICMI15.pdf

Project:http://www.openu.ac.il/home/hassner/projects/cnn_emotions/包含code

预处理：propose novel transformations of imageintensities to 3D spaces, designed to be invariant to monotonic photometrictransformations

 

 

 

ConstrainedConvolutional Neural Networks for Weakly Supervised Segmentation

 

WeaklySupervised Segmentation 比较好的一篇文章，该方法与faster rcnn结合应该是蛮有意思的一篇paper

 

 

LearningDeep Convolutional Neural Networks for Places2 Scene Recognition

ILSVRC 2015 Scene Classification Challenge first place

基于VGG-19 与 GoogLeNet 修改后的两个单独模型，然后合并模型得到第一名

基本idea：

 

 

 

 

 

Poseidon:A System Architecture for Efficient GPU-based Deep Learning on MultipleMachines

 

这是EricP. Xing的第二篇关于分布式训练的文章，这篇文章的key idea有两点还是蛮细致、蛮具有复用性的：

1. dwbp：每个layer bp完了都可以独自先通信

2. sacp：传输整个更新矩阵可能会很大，这样的话有时候可以拆分成两个vector进行传输

 

 

 

NeuralSelf Talk: Image Understanding via Continuous Questioning and Answering

蛮有意思的image自问自答系统，可以往主动学习之类的靠

 

 

DoLess and Achieve More: Training CNNs for Action Recognition Utilizing ActionImages from the Web

利用爬取的数据帮助提升动作识别精确度

 

 

RecentAdvances in Convolutional Neural Networks

该文章对以下方面内容做了摘要

1. CNN网络结构：Convolutionallayer, Pooling layer, Activation function, Lossfunction, Regularization, Optimization；

2. CNN加速与压缩：FFT, Matrix Factorization, Vectorquantization;

3. CNN的应用：Image Classification, Object tracking, Poseestimation, Text detection and recognition, Visual saliency detection, Actionrecognition, Scene Labeling；

 

 

 

BeyondClassification: Latent User Interests Profiling from Visual Contents Analysis

文章做了从image到用户兴趣点的建模