Preamble.

Can a group of very stupid entities come together and do something smart? Fortunately, the answer to this question is a resounding Yes. Case in point: your brain. Each neuron is radically dumb, but in concert with about 100 billion dumb colleagues, our brains are capable of brilliance.

But how does it all work? 100 or 1000 billion cells are each connected to as many as a thousand other cells, and the cells and the connections are unbelievably intricate. There is no single system known to humankind as complicated as the brain. As a result, brain science is still an uncharted scientific frontier. So far, few theories about how brains think have lasted more than about ten years. And the best theories of the present may well be obsolete in a few years. This presents students of brain science -- that’s YOU -- with the opportunity to be among the first to really figure it out.

But how shall we begin? In this lab, we will continue with the widget strategy, but with a difference. Instead of tediously specifying a function for each widget, and working out its exact connections to other widgets, and then figuring out what the whole can do, instead of all that we will get at the thinking mind through very simple generic brains, with lots of widgets but a uniform circuitry that we need not specify in advance. Generic brains share with biological brains the basics. They are made of:

processing units, like biological neurons, which are

connected to one another, as with biological axons and dendrites, and

the connections can vary in strength, just as synapses differ, and

the connections change as a result of learning, resembling "synaptic plasticity."

Generic brains, like real brains, are especially good at cognitive tasks with these features:

Tonight we will examine basic capabilities of generic brains using a simulated connectionist network called Letter Learner. This program is part of a very nice Hypercard stack which will review the basics of neural networks and allow you to play with one.

Follow these steps to load Letterlearner:

1. Open Macintosh HD/Network Servers/Course software.

2. Drag the "Letter Learner" folder to your desktop.

3. Open the folder you have just dragged.

4. Click on the "LetterLearner" stack. The program should launch.

You should begin by clicking on the Introduction. This will teach you the basic operation of the program. From the home card you can also read ?Tutorial and Help. Two of the Tutorial files are essential to master: What is a neural network? and How do neural networks work? You can tour these tutorials at any point in the lab, but you will have to read and understand them in order to complete your report.

At first, you may find it helpful to turn on the Autohelp switch. Turn it off when you are familiar with the operation of the program.

To help understand how letter learner works, you may also find it helpful to turn on the Grid and Show vectors switches on the letter learner main screen.

At any point, you can turn to the questions of your report, but in order to answer them you should spend time both with the tutorial stacks and simply playing around with the program.

You can save time and increase your learning by working with a partner on this. You might discuss your response to each question, but divide the labor of writing. Just make sure that both your names are on any section that either of you hand in.

2. One goal of computer designers has been to make a computer that can read handwriting. (The Apple "Newton" Message Pad is an attempt to create such a computer.) Current handwriting-reading programs in fact work quite a lot like Letterlearner. As you have seen, Letterlearner learns your own personal handwriting, but even when it is limited to the task of learning just your lettering style, it still does a poor job of discriminating letters. Imagine that you are on the design team that will try to improve Letterlearner’s abilities at learning and discriminating letters. Make a list of specific design improvements that will enhance the network’s abilities as a letter reader. That is, describe how the network architecture could be modified to yield better results. For each of your suggestions, provide a rationale explaining why you think it might work. Offer as many suggestions as you can. By the way, you may find many inspirations by asking yourself how the human brain accomplishes the same task.

3. Letterlearner is an obvious "toy" -- a program with very little resemblance to anything in human psychology or neurobiology. So what’s the point? That is, in what ways might the study of artificial neural networks be helpful in the understanding of real neural networks like the one in your head.