The human brain is extremely effective at performing pattern recognition, even in the presence of noisy or distorted inputs. Artificial neural networks attempt to imitate the structure of the brain, often with a view to mimicking its success. The binary correlation matrix memory (CMM) is a particular type of neural network that is capable of learning and recalling associations extremely quickly, as well as displaying a high storage capacity and having the ability to generalise from patterns already learned. CMMs have been used as a major component of larger architectures designed to solve a wide range of problems, such as rule chaining, character recognition, or more general pattern recognition. It is clear that the memory requirement of the CMMs will thus have a significant impact on the scalability of such architectures. A domain specific language for binary CMMs is developed, alongside an implementation that uses an efficient storage mechanism which allows memory usage to scale linearly with the number of associations stored. An architecture for rule chaining is then examined in detail, showing that the problem of scalability is indeed settled before identifying and resolving a number of important limitations to its capabilities. Finally an architecture for pattern recognition is investigated, and a memory efficient method to incorporate general invariance into this architecture is presented—this is specifically tested with scale invariance, although the mechanism can be used with other types of invariance such as skew or rotation.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)