convolutional deep belief networks
self - taught learning
random filters
3.1.1
contrastive divergence: 对比分歧
loglikelihood of training data
sparse regularization
RBM 隐层单元被当做下一层的输入
stacking the CRBMs-》hierarchical object-part decompositions
This paper: two-layers of CRBMs
3.1.2 Local convolutional RBM
connect each hidden unit to only a local receptive field in the visible image, as in the CRBM, but remove the parameter tying between weights
for different hidden units.
disadvantages:
1. computationally intractable to scale this model to high resolution images
2. sensitive to local deformations and misalignments
image->overlapping regions
1. learn useful features for a particular location
2. avoid spurious activations of hidden units
energy function
3.1.3
learning features based on LBP(local binary patterns)
3.2 Recognition algorithm
CSML: cosine similarity metric learning:
ITML: information- theoretic metric learning
x->PCA->y->z->SVM
4 Experimental results
learned filters are more robust that random filters
仅仅在第二层引用了local CRBM
总结
本篇文章将CRBM应用在face verification中, 特征选择为在LBP基础上进行训练得到的特征和在pixel 上训练相结合, 层数设置为两层。
至于local CRBM 和CRBM 的区别,还要多看一些文章