预备知识:图像处理的基本方法, rbm 的实现,多重神经网络实现, c++ 编程...

5
预预预预 预预预预预预预预预预 :, RBM 预预预 预预预预预预预预预 ,, C+ + 预预 预预 预预预预预预预 预预 :, RBM 预 Sparse Coding 预预 ICA 预预预预预 预预GPU 预预MPI 预预预预预预BLAS 预预预 routine 预预 预预 Infiniband GPU 预预 预预预预C++ python( 预预预预 )

Upload: zandra

Post on 05-Jan-2016

652 views

Category:

Documents


0 download

DESCRIPTION

预备知识:图像处理的基本方法, RBM 的实现,多重神经网络实现, C++ 编程 算法:卷积神经网络,多层 RBM 及 Sparse Coding 的变种, ICA 的多层网络 工具: GPU 编程, MPI 多机通信控制, BLAS 相关的 routine 硬件:构建 Infiniband GPU 集群 实现语言: C++ , python( 处理脚本 ). 算法相关论文. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: 预备知识:图像处理的基本方法, RBM 的实现,多重神经网络实现, C++ 编程 算法:卷积神经网络,多层 RBM 及 Sparse Coding 的变种, ICA 的多层网络

预备知识:图像处理的基本方法, RBM的实现,多重神经网络实现, C++编程

算法:卷积神经网络,多层 RBM 及 Sparse Coding的变种, ICA的多层网络

工具: GPU编程,MPI多机通信控制, BLAS相关的 routine

硬件:构建 Infiniband GPU集群 实现语言: C++ , python(处理脚本 )

Page 2: 预备知识:图像处理的基本方法, RBM 的实现,多重神经网络实现, C++ 编程 算法:卷积神经网络,多层 RBM 及 Sparse Coding 的变种, ICA 的多层网络

算法相关论文 LeCunLeCun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied Gradient-based learning applied

to document recognitionto document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.(卷积神经网络起源 )

Lee HLee H, Grosse R, Ranganath R, et al. Convolutional deep belief Convolutional deep belief networks for scalable unsupervised learning of hierarchical networks for scalable unsupervised learning of hierarchical representations.representations. Proceedings of the 26th Annual International Conference on Machine Learning. ACM, 2009: 609-616.(卷积玻尔兹曼机,人脸特征提取 )

Krizhevsky AKrizhevsky A. Convolutional deep belief networks on cifar-10Convolutional deep belief networks on cifar-10[J]. Unpublished manuscript, 2010.(卷积玻尔兹曼机的实现细节 )

Le Q V,Le Q V, Karpenko A, Ngiam J, et al. ICA with Reconstruction Cost ICA with Reconstruction Cost for Efficient Overcomplete Feature Learningfor Efficient Overcomplete Feature Learning[C]//NIPS. 2011: 1017-1025.(Google大规模网络中实现的算法基础,上述算法的变种 )

Page 3: 预备知识:图像处理的基本方法, RBM 的实现,多重神经网络实现, C++ 编程 算法:卷积神经网络,多层 RBM 及 Sparse Coding 的变种, ICA 的多层网络

GPU及大规模计算论文 Krizhevsky AKrizhevsky A, Sutskever I, Hinton G E. ImageNet ImageNet

Classification with Deep Convolutional Neural Classification with Deep Convolutional Neural NetworksNetworks[C]//NIPS. 2012, 1(2): 4.(ImageNet大规模数据的尝试及单机多 GPU模式 )

Le Q V.Le Q V. Building high-level features using large scale Building high-level features using large scale unsupervised learningunsupervised learning[C]//Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference on. IEEE, 2013: 8595-8598.(1000台机器集群 ,Google的尝试 )

Coates ACoates A, Huval B, Wang T, et al. Deep learning with COTS Deep learning with COTS HPC systemsHPC systems[C]//Proceedings of The 30th International Conference on Machine Learning. 2013: 1337-1345.(3 台GPU Infiniband集群,实现楼上的结果,我们的目标我们的目标 )

Page 4: 预备知识:图像处理的基本方法, RBM 的实现,多重神经网络实现, C++ 编程 算法:卷积神经网络,多层 RBM 及 Sparse Coding 的变种, ICA 的多层网络

其他工具和软件 MVAPICH2 : MPI软件,集群通信控制 GPU编程: Programming massively parallel processors (2nd

edition)

BLAS:矩阵运算

数据源 CIFAR-10

ImageNet

Kyoto http://dippix.tp.chiba-u.jp/database/index_e.html

Caltech-101

Page 5: 预备知识:图像处理的基本方法, RBM 的实现,多重神经网络实现, C++ 编程 算法:卷积神经网络,多层 RBM 及 Sparse Coding 的变种, ICA 的多层网络

基本规划

(1)实现 Lee文章中卷积 RBM无监督对图片特征提取 (2-3 weeks)

(2)实现 Hinton 对 cifar-10 或 ImageNet对图片分类预测,单机多 GPU(3-4 weeks)

(3)实现 Coates 用 Infiniband GPU集群对大规模图片的无监督训练 (1-2 months)