Producing pattern examples from "mental" images

The WiSARD (Wilkie, Stonham and Aleksander's Recognition Device) weightless neural network model
has its functionality based on the collective response of RAM-based neurons. WiSARD's learning phase
consists on writing at the RAM neurons' positions addressed (typically through a pseudo-random
mapping) by binary training patterns. By counting the frequency of writing accesses at RAM neuron
positions during the learning phase, it is possible to associate the most accessed addresses with
the corresponding input field contents that defined them. The idea of associating this process with the
formation of ''mental'' images is explored in the DRASiW model, a WiSARD extension provided with
the ability of producing pattern examples, or prototypes, derived from learnt categories. This work
demonstrates the equivalence of two ways of generating such prototypes: (i) via frequency counting
and filtering and (ii) via formulating fuzzy rules. Moreover, it is shown, through the exploration of the
MNIST database of handwritten digits as benchmark, how the process of mental images formation can
improve WiSARD's classification skills.