Class 37: Neural Networks

Held Monday, April 10, 2000

Overview

Today we consider a very different approach toward artificial intelligence: neural networks, which attempt to simulate the way we think.

Question 37 for discussion in today's class: Neural networks take a different approach to artificial intelligence. Rather than trying to mimic the way we think we think, they try to mimic physical aspects of the brain. Do you expect this to be a fruitful road to AI? Why or why not?

Question 38 for disucssion in Tuesday's class: How might you use genetic algorithms to produce a form of intelligent program? Do you expect this to be a fruitful road to AI?

Notes

• Sarah found the following resource for AI programs: http://www.geocities.com/CapeCanaveral/1624/
• Jeff found another interesting page at http://www.emsl.pnl.gov:2080/proj/neuron/neural/what.html
• For tomorrow, read Dewdney, Chapter 16 (Genetic Algorithms)
• I was pleasantly surprised by today's responses. Is it simply that I asked the question differently? Is it that you don't need to write algorithms?

Contents

• Simulating the Brain

Summary

Simulating the Brain

• A different approach to AI
• Basic technique (from Perceptrons): Lots of inputs that are then processed and combined.
• We'll look at the ``perceptrons can't do everything'' example in some detail.
• Advanced technique: layers of ``neurons''
  • What does a neuron do? Combines its inputs and generates an output.
• How do we build these things?
  • No one knows how our brains wire themselves.
  • Neural network technique: ``Learning'' by updating various parts based on sample inputs/outputs.
• We'll look at the backpropagation learning strategy in some non-mathematical detail.