Unification using a distributed representation

Abstract

In the neural network and artificial intelligence research communities at the moment there is an ongoing debate about the representational adequacy of neural networks versus that of classical symbolic systems. Supporters of the classical symbolic AI paradigm argue that systems such as neural networks using a distributed representation do not have the compositionality required to perform structure-sensitive logical tasks. Supporters of the neural network paradigm point to the difficulty symbolic systems have in generalising to novel inputs. We have investigated performing variable binding with a distributed representation by using a neural network to perform unification. The successful performance of this task adds further evidence to the assertion that neural networks using distributed representations can be used to perform complex structure-sensitive tasks.