fann_sparse.h

The fann sparse implementation is a backward compatible MIMO neuron implementation for sparse networks. Using the MIMO Neuron defined by theese functions one gets a MIMO neuron that behaves like an old FANN Neuron i.e. it gets its working parameters from the struct fann. This MIMO implementation is a kind of compatibilty layer to allow old FANN using applications to run unmodified. This code is not fast: the new VFANN implementations will be much smarter (specilized) and faster. The code in here uses swithches everywhere there is need to select something like activation functions and algorithms.

Summary

fann_sparse.h	The fann sparse implementation is a backward compatible MIMO neuron implementation for sparse networks.
Functions
fann_sparse_layer_constructor	Allocates room inside the layer for neurons and connections.
fann_sparse_neuron_constructor	Allocates room inside the neuron for the connections.
fann_sparse_neuron_run	Does a forward iteration on the neuron.
fann_sparse_compute_MSE	Compute the error on a MIMO Neuron after forward propagation of a certain input vector i.e.
fann_sparse_neuron_backprop	Train the MIMO neuron: this function backpropagates the error to the previous layer and computes the weight update.
fann_sparse_neuron_update	the weight update stored in the neuron is applied.
fann_sparse_neuron_private_data	The structure where the connection mask is tored.

fann_sparse_neuron_constructor

FANN_EXTERNAL int FANN_API fann_sparse_neuron_constructor(
    struct fann * ann,
    struct fann_layer * layer,
    struct fann_neuron * neuron,
    struct fann_neuron_descr * descr
)

Allocates room inside the neuron for the connections. Creates a random sparsely connected neuron acconrding to the connection rate specified in descr->private_data as a float value.

fann_sparse_compute_MSE

Compute the error on a MIMO Neuron after forward propagation of a certain input vector i.e. after neuron->run(). The error is the sum, over all the outputs, of the squared difference between the computed output and the desired target also increments a counter because MSE is an average of such errors

After calling this function on a neuron the train_errors array is set to

(desired_output

neuron_value) Compute the error at the MIMO neuron output. We can use the generic version code because we keep the masked weights to 0

fann_sparse_neuron_backprop

Train the MIMO neuron: this function backpropagates the error to the previous layer and computes the weight update. The weight update is not applied here see <fann_neuron_update_connected_any_any> We can use the generic version code because we keep the masked weights to 0

fann_sparse_neuron_update

FANN_EXTERNAL void FANN_API fann_sparse_neuron_update(
struct fann * ann,
struct fann_neuron * neuron
)

Apply the training on the MIMO neuron according to the globally selected algorithm

the weight update stored in the neuron is applied. This function is called once per input pattern when online training is carried on and once per epoch when batch algoritms are selected.

fann_sparse.h

Functions

fann_sparse_layer_constructor