multi-layer neuralnetworks do not use the perceptron learning procedure.
Because this type ofguarantee cannot be extended to more complex networks in which the average oftwo good solutions may be a bad solution.
Linear neurons(linearfilters(線性濾波))
The error surface for a linear neuron
horizontalaxis: weights
verticalaxis: error
Logistic Neuron
dy/dz = y(1-y)
Back propagation algorithm
Ways to reduce overfitting
- weight-decay
- weight-sharing
- early-stopping
- model-averaging
- bayesian fitting of neural nets
- dropout
The squared error asa function of the weights is not a quadratic function when we use a logisticunit.