A Representation for Temporal Sequence and Duration in Massively Parallel Networks

One of the major representational problems in massively parallel or connectionist models is the difficulty of representing temporal constraints. Temporal constraints are important and crucial sources of information for event perception in general. This paper describes a novel scheme which provides massively parallel models with the ability to represent and recognize temporal constraints such as sequence and duration by exploiting link to link interactions. This relatively unexplored yet powerful mechanism is used to represent rule-like constraints and behaviors. The temporal sequence of a set of nodes is defined as the constraints or the temporal context in which these nodes should be activated. This representation is quite robust in the sense that it captures subtleties in both the strength and scope (order) of temporal constraints. Duration is also represented using a similar mechanism. The duration of a concept is represented as a memory trace of the activation of this concept. The state of this trace can be used to generate a fuzzy set like classification of the duration.