[DavidOlmsted]

The Memory Prediction framework as outlined in Jeff's book is an inherently parallel information processing architecture. As such its simulation must be more neural than algorithmic yet it need not be realistically neural. A middle level of abstraction does exist in the form of probabilisic multivalued logic neural networks. You can think of it as an extension of Bayesian networks with additional multivalued logic connectives. (an early example is here: http://www.neurocomputing.org/html/multivalued_logic_nns.html

This network form uses analog values that are within some arbitrary bounds (traditionally between 0 and 1) but since a winner-take-all operation at the network output end selects the greatest valued network line as the network answer the value bounds really don't matter.

The connectives are:

SUMMATION - the probability "and" operation (addition).

CONDITIONAL - a multivalued logic operation that is also a bounded subtraction operation where the output value does not fall below 0.

INCLUSIVE OR - the multivalued (fuzzy) logic operation that passes the greatest value among its inputs (the max operation).

AND - the multivalued (fuzzy) logic operation that passes the least value among its inputs (the min operation).

The analog value represents the probability of state but since states also can be mapped to meanings the analog value also represents the truth value for any meaning associated with its network line. For both probability values and truth values to be co-oexist consistently the value must be interpreted in a range from indeterminant (without evidence) to true. This is different from conventional logic in which the truth value range (metric) is interpreted from "false" to "true".

Adaptability comes from using the probability "or" operation which is a multiplication factor (just like neural networks). This "weight" must be incremented or decremented according to conditions set by the network. The most general and powerful way to set the conditions is to use the enable and change paradigm. This is exactly like classical conditioning where one input enables the weight and another one actually changes it. This paradigm is also used to access RAM with its address and change (read or write) cycle.
_________________
Click to go to my site at neurocomputing.org