Refresher: Perceptron Training Algorithm

Algorithm Perceptron; Start with a randomly chosen weight vector w0;

Let k = 1;
while there exist input vectors that are misclassified by wk-1, do

Let ij be a misclassified input vector;

Let xk = class(ij)ij, implying that wk-1xk < 0; Update the weight vector to wk = wk-1 + xk; Increment k; end-while;
September 23, 2010 Neural Networks Lecture 6: Perceptron Learning 1

Another Refresher: Linear Algebra

How can we visualize a straight line defined by an equation such as w0 + w1i1 + w2i2 = 0? One possibility is to determine the points where the line crosses the coordinate axes: i1 = 0 w0 + w2i2 = 0 w2i2 = -w0 i2 = -w0/w2 i2 = 0 w0 + w1i1 = 0 w1i1 = -w0 i1 = -w0/w1 Thus, the line crosses at (0, -w0/w2)T and (-w0/w1, 0)T. If w1 or w2 is 0, it just means that the line is horizontal or vertical, respectively. If w0 is 0, the line hits the origin, and its slope i2/ii is: w1i1 + w2i2 = 0 w2i2 = -w1i1 i2/i1 = -w1/w2
September 23, 2010 Neural Networks Lecture 6: Perceptron Learning 2

Perceptron Learning Example

We would like our perceptron to correctly classify the five 2-dimensional data points below. Let the random initial weight vector w0 = (2, 1, -2)T. i2
3 2 1

Then the dividing line crosses at (0, 1)T and (-2, 0)T.
1 2 3

-1

-3 -2 -1

Let us pick the misclassified point (-2, -1)T for learning:

i = (1, -2, -1)T (include offset 1)

-1 -2 -3

i1

x1 = (-1)(1, -2, -1)T (i is in class -1)

x1 = (-1, 2, 1)T
Neural Networks Lecture 6: Perceptron Learning 3

class -1 class 1
September 23, 2010

Perceptron Learning Example

w1 = w0 + x1 (let us set = 1 for simplicity) w1 = (2, 1, -2)T + (-1, 2, 1)T = (1, 3, -1)T

The new dividing line crosses at (0, 1)T and (-1/3, 0)T.
i2
3 2 1

-1

Let us pick the next misclassified point (0, 2)T for learning:
1 2 3

-3 -2 -1

-1 -2 -3

i1

i = (1, 0, 2)T (include offset 1) x2 = (1, 0, 2)T (i is in class 1)

class -1 class 1
September 23, 2010 Neural Networks Lecture 6: Perceptron Learning 4

Perceptron Learning Example

w2 = w1 + x2 w2 = (1, 3, -1)T + (1, 0, 2)T = (2, 3, 1)T

Now the line crosses at (0, -2)T and (-2/3, 0)T.

i2
-3 -2 -1

3 2 1

1
1 2 3

With this weight vector, the perceptron achieves perfect classification! The learning process terminates.

-1
class -1 class 1

-1 -2 -3

i1

In most cases, many more iterations are necessary than in this example.
Neural Networks Lecture 6: Perceptron Learning 5

September 23, 2010

Perceptron Learning Results

We proved that the perceptron learning algorithm is guaranteed to find a solution to a classification problem if it is linearly separable.

But are those solutions optimal?

One of the reasons why we are interested in neural networks is that they are able to generalize, i.e., give plausible output for new (untrained) inputs. How well does a perceptron deal with new inputs?

September 23, 2010

Neural Networks Lecture 6: Perceptron Learning

Perceptron Learning Results

Perfect classification of training samples, but may not generalize well to new (untrained) samples.

September 23, 2010

Neural Networks Lecture 6: Perceptron Learning

Perceptron Learning Results

This function is likely to perform better classification on new samples.

September 23, 2010

Neural Networks Lecture 6: Perceptron Learning

Adalines
Idea behind adaptive linear elements (Adalines): Compute a continuous, differentiable error function between net input and desired output (before applying threshold function). For example, compute the mean squared error (MSE) between every training vector and its class (1 or -1).

Then find those weights for which the error is minimal.

With a differential error function, we can use the gradient descent technique to find this absolute minimum in the error function.

September 23, 2010

Neural Networks Lecture 6: Perceptron Learning

Gradient Descent
Gradient descent is a very common technique to find the absolute minimum of a function. It is especially useful for high-dimensional functions. We will use it to iteratively minimizes the networks (or neurons) error by finding the gradient of the error surface in weight-space and adjusting the weights in the opposite direction.

September 23, 2010

Neural Networks Lecture 6: Perceptron Learning

10

Gradient Descent
Gradient-descent example: Finding the absolute minimum of a one-dimensional error function f(x):
f(x) slope: f(x0)

x0

x1 = x0 - f(x0)

Repeat this iteratively until for some xi, f(xi) is sufficiently close to 0.
September 23, 2010 Neural Networks Lecture 6: Perceptron Learning 11

Gradient Descent
Gradients of two-dimensional functions:

The two-dimensional function in the left diagram is represented by contour lines in the right diagram, where arrows indicate the gradient of the function at different locations. Obviously, the gradient is always pointing in the direction of the steepest increase of the function. In order to find the functions minimum, we should always move against the gradient.
September 23, 2010 Neural Networks Lecture 6: Perceptron Learning 12