Page 95 - Anatomy of a Robot

03_200256_CH03/Bergren 4/17/03 12:27 PM Page 80
80 CHAPTER THREE

Synapse

FIGURE 3-3 Human neurons communicating across synapses

are exposed to a series of situations and gradually learn how to deal with them. Neural-
network computers are generally designed with individual “neurons” that can commu-
nicate with one another, especially within their immediate vicinity. They are arranged
in rows and banks of neurons; an example is shown in Figure 3-4.
The results of each layer are fed into a series of communication units that perform
calculations and reroute information to other neurons. The flow of information is shown
in Figure 3-4. A series of real-world events is fed into the inputs at the top; the neural
net processes the inputs and generates responses out the bottom. The results are scored
(by an experienced person) and the score is fed back into the neural network at the top.
The network then readjusts its communication units so it will do better next time.
Certainly, the network will do better the next time it sees the very same events fed into
its inputs. But oddly enough, it often does better on new events it has never seen at its
inputs before. As such, it is learning.
Neural networks can be built in many ways. One researcher took a silicon substrate
(a slab used to build computer chips), hollowed out pits in the substrate, put neurons
into the pits, and allowed the neurons to communicate by connecting synapses.
Computer circuitry was etched in other areas of the substrate. The entire circuit ran on
a combination of glucose and electricity.
Neural networks can be built from hardware (using computer chips) or they can be
simulated in software. There have been many successful applications of neural network