8.1.0.7.5. net-c
(packages/gblearn2/net-c.lsh)

Author(s): Yann LeCun

a neural net class with one convolutional layer and one fully connected layer. The main purpose of this class is to make replicable fully connected networks. Unlike with net-ff, the input spatial dimensions are treated correctly here. So if you want to recognize characters, use this class rather than net-ff.

8.1.0.7.5.0. (new net-c ini inj ki0 kj0 inthick outthick prm)
[CLASS] (packages/gblearn2/net-c.lsh)


makes a new net-cscscf module. ini inj : expected max size of input for preallocation of internal states ki0 kj0 : kernel size for first convolutional layer a standard fully-connected network can be obtained when ini = ki0 and inj = kj0 . inthick : number of input slices outthick : number of outputs. prm an idx1-ddparam in which the parameters will be allocated.