Self-Organized Synfire Chain Formation in Recurrent Networks
Asohan Amarasingham, Brown University
We address a very simple question here: can a sparse and recurrently-connected network of (very simplified) neuron-like elements learn, in an unsupervised fashion, to generate a reliable temporal sequence of activity patterns in response to a weak externally-originated input? The structure of this problem arises in several contexts related to the general theory of learning in neural networks, the biology of neural computation, and the cognitive phenomenon of trace classical conditioning. An approximate parameterization relating the reliability, temporal length, and spatial organization of the generated patterns to the architectural parameters of model networks is developed (via numerical simulations and qualitative analysis), as are the implications of these relationships for the surrounding scientific contexts.
Abstract Author(s): Asohan Amarasingham