The importance of noise for segmentation and binding in dynamical neural systems

Segmentation and binding are cognitive operations that underlie the process of perception. They can be understood as taking place in the temporal domain. In models of nonlinear oscillatory neurons and neuronal cell assemblies, we represent binding by phase locking of assemblies in different networks, and segmentation by phase separation of assemblies in the same network, leading to waveforms of staggered oscillations. Both processes can be facilitated if the inputs to the system, representing simultaneously activated memories, possess noisy components. In the binding problem they serve as a tagging device, driving phase locking between assemblies that belong to different networks but carry the same inputs. In the segmentation problem they allow us to overcome an inherent limitation on segmentation that, otherwise, cannot accommodate more than a few commonly excited memories.