Two-Layer Perceptrons

• Three-layer directed, weighted graph
  • First layer: input nodes
  • Second layer: hidden nodes
  • Third layer: output nodes
• Alternatively, view as two connected perceptrons
  • Output layer of the first perceptron is the input layer for the second
• Computing an output
  • Compute outputs for first perceptron
  • Using these outputs, compute outputs for second perceptron
  • Backpropagation of errors during training
    • First, train the second perceptron
    • Next, calculate errors for each hidden node
      • For each output node
        • Add to the total error the product of:
          • the weight of the edge from the hidden node to the output node
          • the gradient of the output of the input-to-hidden perceptron
          • the error from the hidden to the output node
      • Modify each incoming weight as with the perceptrons

  ---