A self-organizing scale-sensitive neural network

Three adaptive rules are combined in a neural network simulation: an anti-Hebbian inhibitory learning rule, a variant of a Hebbian excitatory learning rule, and a Weber law neuron-growth rule. The neuron-growth rule permits the network to learn and classify input patterns despite variations in their spatial scale. The inhibitory learning rule permits superposition of multiple simultaneous neural activations (multiple winners) under strictly regulated circumstances, instead of forcing winner-take-all pattern classifications. The selected multiple activations can represent uncertainty or multiplicity in perception and pattern recognition. Such behavior is useful in representing visual ambiguity in the aperture problem, determining visual depth from motion parallax, forming a category-recognition network in a complex environment, and representing visual transparency. The techniques illustrate how higher-level pattern sensitivities may self-organize in complex perceptual environments in the presence of multiple scales, multiple patterns, and uncertainty