4

I need some help with a single layered perceptron with multiple classes.

What I need to do is classify a dataset with three different classes, by now I just learnt how to do it with two classes, so I have no really a good clue how to do it with three.

The dataset have three different classes: Iris-setosa, Iris-versicolor and Iris-versicolor.

The url with the dataset and the information is in : http://ftp.ics.uci.edu/pub/machine-learning-databases/iris/iris.data.

I really appreciate any help anyone can give to me.

Thanks a lot!

Santiago Arango Toro
  • 41
  • 1
  • 2

3 Answers3

5

The standard way to do this is called 'one versus all'... you train three perceptrons. First 1 target = is class a?, 2nd perceptron target = is class b? 3rd = is class c. You just train each perceptron separately, and then take max of the three perceptrons to decide class

seanv507
  • 751
  • 3
  • 12
2

Well , When it comes to AI I am an absolute beginner but here is my answer to your question based on my understandings :

a perceptron has only one activation function, therefore it can return only the values of true and false (in most cases true=0 and false=1), so because of that, I don't think that you will be able to accomplish your goal using only one perceptron but you can absolutely do it using multiple perceptrons which essentially is a neural networks, of course training the network would be a lot harder than calculating the changes of weights as you do in perceptrons, You are gonna have to take advantage of a training algorithm such as backpropagation and a sigmoid activation function. I hope my answer was helpful.

Ashkan
  • 93
  • 3
2

Perceptrons, strictly speaking, are binary classifiers.

To make a multi-class classifier, you should switch to a standard feed-forward neural net with a softmax output layer. Without any hidden layers this is equivalent to multinomial logistic regression.

You can also do the one-vs-all trick as @seanv507 suggests. This often works well in practice but there's no strong basis for it in theory, and the true multi-class version is easier conceptually and practically in this case.

Ben Allison
  • 189
  • 1
  • 2