• pattern matching, not sequential logic
• imprecise, probabilistic matches
• inherently parallel, many independent processes
• concept table, internal patterns to match against

   Before we get into the discussion any further, let me give a abbreviated description of a neural network. Neural nets are composed of a multitude of nodes. Each neural node is a very simple computing device. Each neural node can accept inputs, sum them, normalize its output signal, and deliver it to other neural nodes. In the reverse direction, the neural node can accept feedback (how closely its output matches potential outputs) from its output nodes and adjust its linkages to reflect the feedback.

   Pattern matching is at the core of the neural network operation. When inputs to the neural net is received, they are processed through the net. On the output side, the resultant is compared to a set of pre-known concepts. The concept with the closest match is selected as the internal representation of the input the neural net has received.
   There is no overall computer program coordinating the effort of all the various nodes. There is no IF-THEN-ELSE logic determining whether or not the patterns match.    Programming in neural nets are concerned with how fast the net will learn to recognize its inputs and with selection of the pre-existing concepts that it will match the inputs against.
   Neural networks are imprecise in their pattern matching. To take a familiar example, consider a friend's face. It doesn't always look the same, yet you recognize your friend despite minor changes in physical appearance. This imprecision to detail is both the bane and the boon of neural nets. While it allows our recognition of familiar objects in different settings, it also causes the net not to be able to assert with 100% certainty that the input is exactly the same as the pattern it matched against.
   Neural nets are probabilistic in their assessment of