Beyond the Original PNN Model—Kernel Memory for Modeling Various Neural Pattern Processing Mechanisms

In this chapter, the probabilistic neural network, proposed initially by Specht for pattern classification, is extended from various aspects to provide a novel connectionist concept of kernel memory. Kernel memory deals with multiple different domain data in a single network. The units responsible for different domain data can be mutually connected via the lateral connections of the so-called link-weights. The link-weights can have directed connections between the units, enabling flexible network designing and intricate data processing and thus being exploited to model various cognitive faculties of the artificial mind system. Moreover, inherited from the beneficial properties of the probabilistic neural network model, the networks designed within the kernel memory context are transparent, i.e., the analysis of tracing the data flow within the network structure is straightforwardly performed. The localist approach also enables performing incremental learning tasks without the utility of an iterative parameter optimization algorithm, as required in conventional multilayer perceptron neural network models. Although the notion of learning in the context of conventional artificial neural network models is usually confined to iterative network parameter optimization aiming for a particular limited task, it is altered within the kernel memory context under the extended Hebbian principle of allocating new units and establishing mutual connections.