Neural Networks Chapter 5 Joost N. Kok Universiteit Leiden
Simple Perceptrons Learning = find weights by successive improvement from an arbitrary starting point Supervised learning = learning with a teacher Training set = list of correct input-output pairs
Simple Perceptrons Layered feedforward networks Perceptrons (Rosenblatt 1962) Input units Hidden units Output units
Simple Perceptrons Activation function g Thresholds: fix
Simple Perceptron We want actual output pattern to be equal to target pattern:
Simple Perceptron Hetero-association vs. Auto-association Simple Perceptron = one layer perceptron First consider deterministic threshold units Weight vectors
Simple Perceptrons
Simple Perceptrons Desired:
Simple Perceptrons Problem is solvable by simple perceptrons if the problem is linearly separable No thresholds = separating plane goes through origin XOR problem = Boolean exclusive OR
Simple Perceptrons
Simple Perceptrons Learning rule for Simple Perceptron
Simple Perceptrons
Simple Perceptrons
Simple Perceptrons Linear Units
Simple Perceptrons Gradient Descent Learning
Simple Perceptrons Gradient descent algorithm
Simple Perceptrons Nonlinear Units
Simple Perceptrons