8.1.0.7.20. gb-conx
|
(packages/gblearn2/gb-conx.lsh) |
Author(s): Yann LeCun
A module architecture that connects inputs to outputs (with exactly one
input per output but maybe more than one output per input, i.e. a one to
many map). The inputs and outputs are bidimensional (i.e. sequences of
vectors). It has no adjustable parameters. Instead it has internal
non-adjustable parameters specifying the input/output connectivity.
8.1.0.7.20.0. (new gb-conx n-outputs n-inputs indices)
|
[CLASS] (packages/gblearn2/gb-conx.lsh) |
Construct a gb-conx architecture. The parameters are given when
constructing the corresponding machine.
See: machine
See: gb-module
See: gb-conx
8.1.0.7.20.1. (==> gb-conx fprop parameters input output)
|
[MSG] (packages/gblearn2/gb-conx.lsh) |
Fprop method for a gb-conx architecture. Just gather the inputs into the
output vectors. Both the input and output are assumed to be sequences of
vectors [i,t] (2D matrices).
See: gb-conx
8.1.0.7.20.2. (==> gb-conx bprop input output)
|
[MSG] (packages/gblearn2/gb-conx.lsh) |
Bprop method for a gb-conx architecture. Just scatter with accumulate
the derivatives of the outputs into the derivative of the inputs. Both
the input and output are assumed to be sequences of vectors [i,t] (2D
matrices).
See: gb-conx
8.1.0.7.20.3. (==> gb-conx generate input output)
|
[MSG] (packages/gblearn2/gb-conx.lsh) |
:output:x is a matrix whose columns have zeros and a single one.
According to the permutation matrix parameters
, a similar matrix :input:x is produced.
8.1.0.7.20.4. sum-arch
|
(packages/gblearn2/gb-conx.lsh) |
An architecture that takes the sum of its inputs. The input is a 2D
matrix whose first index is n-inputs
and second index is variable (e.g. time). The output is zero-dimensional
(an idx0). There are no parameters.
See: gb-module
See: (new sum-arch)
See: (==> sum-arch fprop
parameters input
output )
See: (==> sum-arch bprop
parameters input
output )
See: (==> sum-arch
output-dim inputx )
See: (==> sum-arch
generate parameters
input output )
8.1.0.7.20.5. (new sum-arch)
|
[CLASS] (packages/gblearn2/gb-conx.lsh) |
Construct a sum-arch gb-module.
See: machine
See: gb-module
See: sum-arch
8.1.0.7.20.6. (==> sum-arch output-dim inputx)
|
[MSG] (packages/gblearn2/gb-conx.lsh) |
8.1.0.7.20.7. (==> sum-arch fprop parameters input output)
|
[MSG] (packages/gblearn2/gb-conx.lsh) |
Fprop method for a sum-arch architecture. Just add all the inputs. The
input is assumed to be any matrix (except an idx0). The output is an
idx0.
See: sum-arch
8.1.0.7.20.8. (==> sum-arch bprop parameters input output)
|
[MSG] (packages/gblearn2/gb-conx.lsh) |
Bprop method for a sum-arch architecture.
See: prod
8.1.0.7.20.9. (==> sum-arch generate parameters input output)
|
[MSG] (packages/gblearn2/gb-conx.lsh) |
Generate a 2-d sequence of inputs. A prior distribution on sequence
length is chosen here to be uniform in [5, 6, ..10]. Ignore the output.
For each time step, the input vector is a vector of 0's and a single 1,
chosen uniformly.