G-RAM is a branch of study in neural networks and artificial intelligence. A neural network is a simulation of a group of neurons. For simplification purposes, in this paper, binary inputs and outputs are considered in a neural network. RAM means that, as in computer memory, the inputs and the one output of a selected network are stored in a computer record (random access memory). For example, in “0110110/1,” the input is to the left of the slash, and the output (1) is to the right. To find a match, do a search of RAM memory on input. The binary values could be generalized beyond binary values to the input and the output or response blackboard.

The G in G-RAM means generalized. That is, develop a set of inputs and do not find a match. Now, all is not lost. Find the two nearest neighbors. Treat the inputs as a binary number. Take the difference between the new input value and the neighbor inputs. The smaller difference is called the Hamming distance. Interpolate and use the nearest neighbor output as the output value for the new input.

The initial RAM memory load is called the training set. This paper includes a discussion of the performance of the G-RAM model in rejecting noise. The best starting point is a random training set. The example used in the paper is pattern recognition of various types of a facial image, with a training set of ten images of facial expressions and lighting. Future work could be on enhancing the performance of the G-RAM.