(Line + Point) attractors in recurrent networks
Gatsby Computational Neuroscience Unit, UCL , UK
Point attractor dynamics have been proposed as the underlying mechanism for performing tasks such as objects recognition. In the models that incorporate this view, a point attractor in the phase space of the system is associated to each object and convergence to one such point is regraded as recognition of the associated object. However, it is more likely that the brain does not store information about objects in the form of point attractors but rather more complicated continuous manifolds. In this view, a higher dimensional manifold is associated to each object; which manifold the network is on represents the identity of the object and the position on the manifold represents some continuous variable such as position or view of the object. Considering a network with spatially organised connectivity, I show how storing and retrieving such manifolds is possible in a recurrent network.