|
Hi guys
I read the paper Convolutional Deep Belief Networks for Scalable Unsupervised Learning of Hierarchical Representations.In this paper,it propose that every weight has k groups. For example,if the we choose 69x69 area as the input data,the size of weight is 10x10x24 ,where 24 stands for k groups. And after convolution we get 60x60x24 output. And i want to train the 2layer.What is the input of the 2layer?If 60x60x24 as the input, the dimension will larger than original input data. |
|
Look at it this way- 1st layer input is 69x69x1, there is one input channel (grayscale). 2nd layer input is 60x60x24, there are 24 input channels. So, yes, now you are using 24 channels to represent your input data. That is, I think, what it is meant by over representation by k times as mentioned in the paper. Thank you very much to answer my question. I understand that your means is in 2nd layer every channel 60x60 as a input. After computing , the results add together.It is right?
(Jun 25 '14 at 21:56)
gjy
yes, each pixel is calculated by w1x1 + w2x2 +w3x3 +... + b, where wn is the filter for input channel-n, convolution with the input xn.
(Jun 26 '14 at 07:01)
Ng0323
|
