Pattern segmentation in a binary/analog world: unsupervised learning versus memory storing.
We discuss the problem of segmentation in pattern recognition. We adopt the model and the general approach in the landmark paper by Wang, Buhmann and von der Malsburg (Neural Computation, (1990), 2, 94-106), and expand their model in a number of ways. We review their solution to the segmentation problem in associative memory, which consists in feature binding being expressed by synchrony relations between oscillators or populations of neurons. We extend the model by introducing a law of synaptic change, which allows the network to learn by structuring itself in response to stimuli with relevant features. We discuss the problem of interference between pattern completion and the learning of new memories. We also propose a form of multiplexing of input information taking advantage of the time-structure of the neurons' response. It is based on the assessment of analog as well as of binary properties of the stimuli and provides for an enhancement of the network's processing capacity. The relevance of the results for biological systems is pointed out.