In this paper we develop a model which, through a dynamic modulatory local feedback mechanism, is capable of categorizing multiple simultaneously-presented input patterns while only previously being trained on single patterns. The key process is shown to be segmentation where low level feature detectors are grouped with their corresponding high level categories. While most existing approaches use oscillatory dynamics, our model uses a concept of fuzzy membership sets, thus potentially cutting down on computational requirements and improving convergence properties.The model is based on two main principles: top-down modulation of synchrony and resonant increase of amplitudes during partial synchronization. We show that top-down feedback considerably reduces ambiguity in visual scenes with superimposed patterns. This is a task which traditional artificial neural networks (ANNs) are not ideally suited to solve.Our model is based on competitive networks without any lateral inhibition. The competition arises from feature nodes being attracted toward category nodes based on current context. We present our model and perform simulations of simplified character recognition to show the dynamics of our model.