Calculus of Networks of Adaptive Elements

Intensive effort over the last four years has led to a variety of achievements in the field of trainable pattern classifying systems. Particular progress has been made in the construction, application, and theory of learning machines – trainable systems are often consist of networks of adaptive threshold logic units. The following developments are illustrative of some of the accomplishments of the learning machine research program at the Stanford research institute:

• Development of a low cost high speed electronically adjustable weighing elements
• Design and construction of a large scale learning machine
• Application of learned learning machine techniques to certain weather prediction problems
• Development of techniques for determining structural features of patterns
• Development of techniques for clustering similar pattern
• Contributions to the theory of the trainability and capacity of a threshold logic unit
• Investigations into the mathematical theory of networks of threshold logic units.

The mathematical results obtained so far represent the beginnings of a theoretical base for understanding of the mini proposed trainable systems. The ultimate goal of the current our ADC contract is to provide such a base by developing a calculus of networks of adaptive elements. The major steps that have been taken towards this goal are reported in detail in Rath $.15 and the present state of this research can be summarized as follows…