Perceptron

#machine-learning #interview

Perceptron is a linear classifier
- So it is used only when data are linearly separable
Initially unit step activation function was used
In short, it's a simple model of $y = w x + b$ where
- prediction = 1 if y > 0
- otherwise, prediction = 0

Steps

For d=2 dimensional data,
1. W = ( $w_{0}, w_{1}, w_{2}$ ) - (d+1) dimension
2. X = ( $x_{0}, x_{1}, x_{2}$ ) - (d+1) dimension
3. Here $x_{0}$ is always $1$
Now calculation $\hat{y} = w_{0} + w_{1} x_{1} + w_{2} x_{2}$
1. if $\hat{y} > 0$ , then upper side
2. else, lower side
Update:
1. For all incorrectly classified ones:
$w_{n e w} = w_{n e w} + η d x$
1. d = 1, if it needs to be on upper side
2. d = 0, if it needs to be on lower side

Related Notes