Image Segmentation Using Neural Oscillators
Konkle, Jiang, Zhang, Gurel, Scheper, & Craciun
Synchrony in neural oscillations is an observed phenomenon in neurobiology and has been proposed as a possible neural mechanism addressing the binding problem (Malsburg and Buhmann, 1992). We attempted to mathematically model this binding synchrony as it pertained to a twodimensional binary visual input more simply, neurons corresponding to an object A should fire together, and these neurons should fire out of phase with neurons corresponding to object B. We used reduced HodgkinHuxley equations for our neuronal model, eliminating the spatial dimension in order to work with ordinary differential equations. We added ‘synchrony terms’ to these equations accounting for the synaptic connectivity, changing the strength and type of these connections in order to achieve object binding and image segmentation. Using nearest neighbor excitatory connections in the sensory (input) neuronal layer, and introducing a global inhibitor receiving from and transmitting to every sensory cell (Terman and Wang, 1995), we found that this connectivity would produce the desired results when a proper balance of the synaptic conductance parameters was found. We briefly examined further extensions to this model by incorporating moving objects.
Konkle, T., Jiang, N., Zhang, J., Gurel, F., Scheper, C., & Craciun, G. (2004). Image segmentation using neural oscillators. Technical Report No. 26. The Ohio State University: Mathematical Biosciences Institute Technical Report Series. Online.