UC Riverside

Psychological research shows that in order to visually identify an object, our brain generates an initial hypothesis of what it is, and tests stored patterns from memory against this hypothesis to see if the expectations or predictions that are based on this hypothesis are true. If they are true, then the hypothesis is correct, if not, the hypothesis is not supported and the cycle repeats until it reaches a conclusion on what it sees. In other words, “we see what we expect to see”. In this thesis, we show how a neural network based pattern recognition system built upon this psychological concept can help improve the success rate of visual pattern recognition using offline handwritten character recognition as an example. Where an initial neural network classification of an n letter word is separated using a threshold and refined over multiple iterations of a reasoning and recognition process using specifically chosen weighted letter mask templates applied on every new classification. Statistical knowledge guides what threshold to use and what weights to use for the masks while symbolic knowledge guides which letter mask templates to use.