神经网络课4-123教案 Chapter 4 Multiple Feedforward Neural Network 研究生课件教材课程.ppt

Chapter 4 Multiple Feedforward Neural Network;;Neuron j ?wijxi;;Assuming:;wij,new= wij,old+?wij;NN with thresholds;Calculating steps of BP algorithm;An illustrative example: fault diagnosis;Specific conditions:;A three-layer perceptron neural network;Attempts to properly map given inputs with desired outputs by minimizing an error function.;Step 1:;Step 2:;Note that had we not normalized our input values to the ranger [0,1], we would have the following results:;Step 3:;Step 4:;Step 5:;Step 6:;Step 7:;A comparison of the old and new weight factors shows:;Step 8:;Step 9:;Step 10:;Step 11:;4.2 Improving on BP Algorithm;+ momentum;New weight factor;To see the comparison, we will apply the delta rule for the three-layer-feedforward perceptron to the fault-diagnosis problem of section 4.1.;Step 2:;Step 3:;Step 5:;Step 6:;Example:;Step 7:;Example:;Step 8:;Example:;We also calculate the weight change ?wjk,new using the same learning rate (?=0.9) and momentum coefficient (?=0.7):;Step 9:;Example:;Number of time steps: 3860 Desired values: d1=1 d2=0 d3=0 Actual values: c1=0.9900 c2=0.0156 c3=0.0098 Percent error: e1=1.00% e2=1.56% e3=0.98%;C. Alternative Formulation of Backpropagation Learning; Output from the hidden layer : ;An alternative architecture of the three-layer feedforward network;This bias input a0 is fed to all nodes in the hidden and output layers with weight factors voj and wok (j=1 to 3;k=1 to 3),and the resulting weighted biases replace the internal thresholds T2j and T3k (j=1 to 3;k=1 to 3).;大作业：;4.2.2 overview of special modified methods;u=1.1~1.3, d=0.7~0.9;See page 79~92;New;B. Activation functions;4.3 Practical aspects of neural computing;;C. Normalizing input and output data sets;Method 1;Method 2;Method 3;;E. Setting the learning rate and momentum coefficient;With a low learning rate, training may be trapped in a locaal minimum;Adding momentum term;;F. Selecting the proper transfer func