jbn@glacier.UUCP (12/05/87)
I've been implementing Rumelhart's learning technique, and
observing how fast it learns, what factors affect the learning rate,
and how my results compare with his. Suddenly it struck me - I'm
repeating someone else's experiment, and comparing my data with his.
It's rare in this field to be able to repeat the experiments of
another and actually compare numerical results. In this area, we
can do it. We can conduct repeatable experiments and objectively
validate the work of others. This is real science. Instead of
arguing, we converge on accepted, repeatable results. The
scientific method works here.
It's interesting that in the area of AI where things seem most complex,
chaotic, and noisy, one can do good experimental science. This field
may move forward rapidly.
John Nagle