Talk:Perceptron: Difference between revisions

== Tasks ==

Excellent: I was looking for some machine-learning style tasks, and found this one!

As it's still in draft, could I suggest some additional tasks to be done? I also suggest the image output be moved to 'extra credit': image output is difficult for languages without a natural gui library, and there also seem to be problems uploading images to the site, so results are hard to show. (Although I do like the Java version's dynamic graphical output.)

As a core set of tasks, how about simply showing the perceptron being trained and working:

# Create a set of training data of at least 5000 random points, classifying the points with a line of your choice.

# Train a perceptron from the training data.

# Show the final perceptron's weights

# Create a new test set of at least 1000 random points, and show the percentage correct achieved by the final perceptron.

As 'extra credit' tasks:

# Draw an image showing the target line and the decisions made by the perceptron on the test set.

# Show how performance on a fixed test set varies with an increasing range of training data, say 1000 to 20000 points in steps of 1000.

I shall add a Scheme example illustrating these core tasks and the second extra one.

-- [[User:Peter]] 17:15, 25 February 2017 (UTC)