Context-Sensitive Weights For A Neural Network

This paper presents a technique for making neural networks context-sensitive by using a symbolic context-management system to manage their weights. Instead of having a very large network that itself must take context into account, our approach uses one or more small networks whose weights are associated with symbolic representations of contexts an agent may encounter. When the context-management system determines what the current context is, it sets the networks' weights appropriately for the context. This paper describes the approach and presents the results of experiments that show that our approach greatly reduces the training time of the networks as well as enhancing their performance.