Neuromorphic vision systems are commonly based upon models of biological neural circuits. Currently, the circuits and processing in the retina are the best understood, which has enabled neuromorphic engineers to implement fairly realistic silicon models of retinal processing. However, as we move towards higher levels of processing in the brain, our knowledge about the neural circuitry decreases dramatically....[ Read more ]

Neuromorphic vision systems are commonly based upon models of biological neural circuits. Currently, the circuits and processing in the retina are the best understood, which has enabled neuromorphic engineers to implement fairly realistic silicon models of retinal processing. However, as we move towards higher levels of processing in the brain, our knowledge about the neural circuitry decreases dramatically.

Given this lack of knowledge, it might seem an impossible task for neuromorphic engineers to design silicon models of neural processing in areas beyond the retina. However, rather than depending on explicit knowledge about neural circuitry, we may be able to wire circuits modelling these areas by exploiting another biological mechanism: self organized development.

We describe an algorithm for self-organizing connections from a source array to a target array of neurons that is inspired by neural growth cone guidance. Each source neuron projects a Gaussian pattern of connections to the target layer. Learning modifies the pattern center location. The small number of parameters required to specify connectivity has enabled this algorithm's implementation in a neuromorphic silicon system. We demonstrate that this algorithm can lead to topographic maps similar to those observed in the visual cortex, and characterize its operation as function maximization, which connects this approach with other models of cortical map formation.