Robust Character Recognition Using a Hierarchical Bayesian Network

There is increasing evidence to suggest that the neocortex of the mammalian brain does not consist of a collection of specialised and dedicated cortical architectures, but instead possesses a fairly uniform, hierarchically organised structure. As Mountcastle has observed [1], this uniformity implies that the same general computational processes are performed across the entire neocortex, even though different regions are known to play different functional roles. Building on this evidence, Hawkins has proposed a top-down model of neocortical operation [2], taking it to be a kind of pattern recognition machine, storing invariant representations of neural input sequences in hierarchical memory structures that both predict sensory input and control behaviour. The first partial proof of concept of Hawkins' model was recently developed using a hierarchically organised Bayesian network that was tested on a simple pattern recognition problem [3]. In the current study we extend Hawkins' work by comparing the performance of a backpropagation neural network with our own implementation of a hierarchical Bayesian network in the well-studied domain of character recognition. The results show that even a simplistic implementation of Hawkins' model can produce recognition rates that exceed a standard neural network approach. Such results create a strong case for the further investigation and development of Hawkins' neocortically-inspired approach to building intelligent systems.