Please enable JavaScript.
Coggle requires JavaScript to display documents.
FFNN - Coggle Diagram
FFNN
Supervised learning
error correction principle
minimization of the error signal
update by iteratively minimizing the difference of the predicted network output and reference
learning = specialize the network to a specific task
any pattern as a reference output
training dataset with output reference data set
Binary Units
Perceprtron Learning Rule
solutions
can not exists
the algorithm could also not converge
corrects the weight vector if and only if a misclassification occurs
any pattern corresponds to direct update of the weights
online update
weight vector orthogonal to the decision boundary
Linear Non-linear Units
Gradient descendent method
iterative optimization alg. approching a local minimum
derivative of the activation function
stop criterion
No hidden layers
Delta rule
online
batch update
with hidden layers
Back propagation
propagation of the error signals from the output layer back to th hidden layers
iteratively compute gradients for each layer
1_ DIRECT COMPUTATION computation of the neuron output layer by layer
update of the weights layer by layer working backwards
Definition
signal go from k-1 to k to k+1
NO
Inter
Intra (self connection)
linking together multiple MCPs
Problems
selection of the activation function
linear
non linear
learning rate n
small values
good approximation
slow convergence
large values
fast convergence
local minima oscillations
supervised learning
consistent training data set
represent a all the classes that should be learned
fitting quality
underfitting
overfitting
convergence
stop criterion
Topology of the network
super distributed information
fully interconnected
distributed information