Consolidation and learning: a connectionist model of human credit assignment

This dissertation concerns the issue of human credit assignment, particularly why some things are so easily remembered while others are forgotten. The point of view taken is an adaptive one: differences in the strength of learning from situation to situation are seen as reflections of the differences in the probable importance of each situation. One difficulty with such an approach is that it appears to imply that humans have the ability to effectively evaluate importance--a scenario which is often unlikely, especially in novel domains. The alternative offered is that the human cognitive architecture is sensitive to a variety of generic situations, and that these sensitivities are automatically reflected in learning. Insight into these sensitivities is provided by an examination of the consolidation process, a brief period of transition during which memory is fixed or made permanent. There is a large set of evidence linking consolidation with the ultimate strength of learning. Consolidation is modelled in this dissertation with TRACE, a connectionist simulation based upon Hebb's cell assembly construct. Learning in TRACE comes as the result of correlated neural activity; factors which influence such activity, therefore, will necessarily impact learning. This provides the human cognitive architecture with the sensitivity required by the learning system. One of the key factors which impacts consolidation is arousal. The arousal level of an organism, in turn, is sensitive to "important" events such as pleasure, pain and confusion. The relationship between arousal and the strength of learning is well known: When arousal is high, learning is strong; when arousal is low, learning is weak. Less well known is the impact of arousal on short term performance. High levels of arousal have been shown to result in "reminiscence," a situation where performance improves over time without intervening practice. In this case high levels of arousal actually impair short term recall. This result is counter-intuitive and until now has never been successfully modelled. Reminiscence can be modelled in TRACE, and within the context of a consolidation-based learning system, the effects can also be shown to be adaptive.