14.1.4 The Bp Trial Process

The BpTrial process is the only Bp-specific process type needed to perform simple feed-forward backprop. The Loop function of this process simply propagates activation forwards through the network, and then propagates the error backwards. This assumes that the layers are ordered sequentially in a feed-forward manner. Note that the process does not actually "loop" over anything, so it has no counter. See section 12.5.1 Presenting a Single Event: TrialProcess for more information on the trial process type.

The following functions are defined on the trial process, each of which performs one step of the backpropagation computations:

Compute_Act()
Goes layer-by-layer and computes the net input and the activation of the units in the layer.
Compute_Error()
Computes the error on the output units, which is only done during testing, and not training.
Compute_dEdA_dEdNet()
Computes the derivative of the error with respect to the activation and then with respect to the net inputs. This goes in reverse order through the layers of the network.
Compute_dWt()
Computes the dEdW for all the weights in the network.