• The perceptron is the simplest form of a neural network used for classification of a special type of patterns Which is linearly separable patterns . • single layer perceptron consists of single neuron With adjustable synaptic weights and threshold as shown in figure. • By expanding the output layer of the perceptron to include more than one neuron , We may form classifications with more than 2 classes but linearly seperable Classes . Single layer perceptron • The resulting sum from the summing node is applied to a hard limiter • So then neuron produces an output equals to +1 if the hard limiter input is positive and -1 if the hard limiter input is negative . •
• The purpose is to classify between two classes C1 and C2
The decision rule is to assign the point represented by the inputs x1 , x2 , … xm to class C1 if the perceptron output is +1, and to class C2 if the perceptron output is -1. The two decision regions are separated by a hyperplane defined by: • If we have only 2 inputs, the decision boundary takes form of straight line as shown. • Above the line the decision is C1 and below the line the decision is class C2 • The effect of the threshold b is to shift the boundary away from the origin. • The weights could be fixed or adapted on an iteration by iteration basis. The Perceptron Convergence Theorem • The threshold b(n) denotes to nth iteration 𝑇 • 𝑋 𝑛 = −1, 𝑥1 𝑛 , 𝑥2 𝑛 , … . . 𝑥𝑚 (𝑛) • 𝑊 𝑛 = 𝑏 𝑛 , 𝑤1 𝑛 , 𝑤2 𝑛 , … . . 𝑤𝑚 (𝑛) 𝑇 • 𝑣 𝑛 = 𝑊𝑇 𝑛 𝑋 𝑛 • WT X ≥ 0 for every input vector X belonging to C1 • WT X < 0 for every input vector X belonging to C 2 • The training problem is to find a weight vector W such that the inequalities above are satisfied. The Perceptron Convergence Theorem • Parameters and variables: • X(n) The input vector • W(n) The weight vector • b(n) The threshold • y(n) the actual response • d(n) the desired response • η learning rate parameter from 0 to 1 The Perceptron Convergence Theorem • Step 1 Initialization: Set W(0)=0, then perform the following computations for n=1,2,… • Step 2 Activation: At time n activate the perceptron by applying x(n) and d(n) • Step 3 Computation of actual response: 𝑦 𝑛 = 𝑠𝑔𝑛 𝑊 𝑇 𝑛 𝑋(𝑛) +1 𝑖𝑓 𝑣 > 0 𝑆𝑔𝑛 𝑣 = −1 𝑖𝑓 𝑣 < 0 • Step 4 Adaptation of weight vector: 𝑤 𝑛+1 =𝑊 𝑛 +𝜂 𝑑 𝑛 −𝑦 𝑛 𝑋 𝑛 +1 𝑖𝑓 𝑋 𝑛 𝑏𝑒𝑙𝑜𝑛𝑔𝑠 𝑡𝑜 𝐶1 𝑑 𝑛 = −1 𝑖𝑓 𝑋 𝑛 𝑏𝑒𝑙𝑜𝑛𝑔𝑠 𝑡𝑜 𝐶2 • Step 5: Increment time n by 1 unit and go back to step 2 • Error signal = d(n)- y(n) Multi-Layer Perceptron Back- Propagation Algorithm Summary
The maximum margin classifier chooses the hyperplane that results in the largest separation, or margin, between the two classes. This hyperplane will best generalize to unseen data points