Learning sequences in an associative network: A step towards cognitive structure.

A crucial step towards the representation of structured, symbolic knowledge in a connectionist system is an associative, neurophysiologically plausible theory of sequence learning. Learning internal representations for sequences is important on adaptive as well as theoretical grounds, since it provides the capability of anticipating probable next states of the world. Through computer simulation of a connectionist, neurocognitive model, this work begins the development of a theory of sequence learning. Sequence learning is characterized as (1) the formation of an internal representation from a series of perceptual events experienced in order over time, and (2) the use of this internal representation to predict events in the absence of supporting environmental input. Development of the model proceeds in a failure-driven fashion, wherein modifications and new mechanisms are introduced to correct specific shortcomings revealed in the simulations. The starting point is a network architecture designed for perceptual learning of compact representations for features and objects. Units of the network do not model single neurons but cell assemblies: multi-neuron groups which exhibit a persistent and complex time course of activity. Unit dynamics are modelled with a system of difference equations which compute state trajectories for three unit state variables, with activation being the primary variable. At any time the content represented by a unit is either latent (no activation), conscious (peak activation), or held in immediate memory (intermediate activation). A modified Hebbian learning rule is introduced which strengthens connection weights in response to consecutive, rather than concurrent, activity in connected units. The importance of controlling the activity generated by positive feedback within and between units is emphasized and specific dynamic control functions are elucidated. Using simulation and conceptual analysis, a set of complimentary control mechanisms, each satisfying a particular functional requirement, is identified. Together these mechanisms provide effective control of activity which permits training and testing the system on sequence learning tasks. A series of simulations then demonstrate elementary learning of single and intersecting sequences. It is concluded that an associative explanation of sequence learning is possible within a neurophysiologically plausible connectionist architecture.PhDComputer Science and EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/105608/1/9135697.pdfDescription of 9135697.pdf : Restricted to UM users only