Generalised brain-state-in-a-box based associative memory for correcting words and images

Human brain has amazing capability to recall the information if a small but sufficient clue is presented. This is known as content based recall or associative memory. Hopfield nets possess capability to recall the patterns based on their content. However, Hopfield net suffers from certain drawbacks which prevent its wide use as a means of realizing associative memory. The most notable among them are very low capacity (0.15N where N is the no. of neurons) and concomitant presence of very large number of spurious states, the undesired stable patterns. In this paper an approach, which is based on generalised brain-state-in-a-box (GBSB) model has been used to correct misspelled English words (strings). Also, an experiment has been conducted for recalling the original image from the pool of stored images when corrupted image is presented as input. Generalised brain-state-in-a-box (GBSB) is used to store the desired patterns as stable states of recurrent neural network. As the capacity of this associative memory is more than that of Hopfield network having the same number of neurons and connections and the number of spurious states are also quite less, this model provides an effective practical approach for associative recall.