Feedforward, feedback neural networks with context-driven segmentation and recognition
Leon N. Cooper, Predrag Neskovic · 1999
It is generally accepted that our experience influences our perception of the outside world. What we expect heavily shapes what we believe we see or hear. Many of the ambiguities, present at the low level of recognition or processing, seem impossible to resolve without taking into account cognitive level expectations. The presence of feedback connections in the human visual system is well documented and for many areas of the brain feedback connections are an order of magnitude more numerous than feedforward connections. Despite these massive feedback pathways, the role of the feedback is still elusive. In this thesis we present a working neural network system that explicitly implements context information to segment, modify and organize bottom up information so as to improve recognition. We start by introducing a new energy function that represents an object as a set of interacting parts. The value of the energy function is related to the confidence of classifying a pattern as one of the memorized objects, and maximization of the energy function leads to pattern segmentation and recognition. The energy function is then mapped into the neural network that consists of a large number of interconnected units, in which each unit performs a simple operation, such as the summation or maximization of its input elements. The neural network finds an approximate solution of the energy function in just a few steps. It relies heavily on feedback connections and has properties that resemble the functioning of the human visual system, like selective attention, saccadic eye motion, and the use of the contextual information to guide recognition and resolve ambiguities on various levels. We present the results of our recognition system applied to on-line cursive script.