The development of a neural network model of associative memory based on a classifier
Description
Neural network models of associative memory offer an alternative to traditional methods of storing and retrieving information. They are more robust, faster, and are capable of forming generalizations of the information to be stored. However, most neural network associative memory models can only store and accurately retrieve a limited number of distinct memories. In addition, retrieval is based solely on the similarity or closeness of the individual attributes of the retrieval cue to those of the stored objects In this work a classifier neural network is combined with a Coulomb energy network to produce a new neural network model. The characteristics of the model do not impose a limitation on the number of memories which can be accurately stored and retrieved. The generalizations formed by the classifier are used to guide the retrieval mechanism, so retrieval is not limited to an a priori set of stored memories. An added benefit derived from the use of the classifier is that the retrieved object can also be made to satisfy an additional constraint of meeting a certain classification requirement, a functionality not addressed by other models The Coulomb energy network transforms the retrieval cue by using gradient descent on an energy function which is minimized when the cue is close to objects of the desired classification. An annealing approach to retrieval is introduced to overcome the problem of local minima associated with gradient descent, resulting in a more robust model. An approach is developed to allow the model to retrieve objects based on arbitrarily complex relationships among several attributes, a functionality not provided elsewhere. This is done by representing these higher-order retrieval criteria as additional constraints to be satisfied and incorporating these constraints into the energy function to be minimized. Implementation strategies which take advantage of the inherent parallelism of neural networks are investigated. The result is a new neural network model of associative memory which is more robust and versatile than other models