11 1 Backpropagation
11 2 Multilayer Perceptron R – S 1 – S 2 – S 3 Network
11 3 Example
11 4 Elementary Decision Boundaries First Subnetwork First Boundary: Second Boundary:
11 5 Elementary Decision Boundaries Third Boundary: Fourth Boundary: Second Subnetwork
11 6 Total Network
11 7 Function Approximation Example Nominal Parameter Values
11 8 Nominal Response
11 9 Parameter Variations
11 10 Multilayer Network
11 Performance Index Training Set Mean Square Error Vector Case Approximate Mean Square Error (Single Sample) Approximate Steepest Descent
11 12 Chain Rule Example Application to Gradient Calculation
11 13 Gradient Calculation Sensitivity Gradient
11 14 Steepest Descent s m F ˆ n m F ˆ n 1 m F ˆ n 2 m F ˆ n S m m = Next Step: Compute the Sensitivities (Backpropagation)
11 15 Jacobian Matrix F Ý m n m f Ý m n 1 m 0 0 0f Ý m n 2 m 0 00 f Ý m n S m m =
11 16 Backpropagation (Sensitivities) The sensitivities are computed by starting at the last layer, and then propagating backwards through the network to the first layer.
11 17 Initialization (Last Layer) a i n i M a i M n i M f M n i M n i M f Ý M n i M === s i M 2t i a i – –f Ý M n i M =
11 18 Summary Forward Propagation Backpropagation Weight Update
11 19 Example: Function Approximation Network + - t a e p
11 20 Network Network a p
11 21 Initial Conditions
11 22 Forward Propagation
11 23 Transfer Function Derivatives
11 24 Backpropagation
11 25 Weight Update
11 26 Choice of Architecture Network i = 1i = 2 i = 4i = 8
11 27 Choice of Network Architecture
11 28 Convergence
11 29 Generalization