Short-term Learning for Long-term Retention : Dynamic Associative Memory
Abstract
Instead of characterizing transfer from short-term memory to long-term memory as the relocation of information from one structural system to another, I propose a theory that conceives of transfer as the learning processes that act on and transform the representations of the information itself. Dynamic Associative Memory posits that recently encoded memories are supported by active maintenance and the relevance of the current context. Over time, the current context becomes less relevant; therefore, the brain must learn contextually invariant associations between memories so that they may support themselves. I instantiated my theory in the ACT-R cognitive architecture and created a new module to automate and fully integrate attentional refreshing into the architecture. The DAM module extends ACT-R's spreading activation to allow activation to be shared among related items in declarative memory. It implements a novel associative learning process based on causal inference that stochastically generates new memory traces for associations between items proportionate to the causal power of one item to predict the other. I also developed another module to provide ACT-R models with a principled method for updating temporal context, and I proposed similarity functions for quantifying the contextually invariant relatedness of hierarchical relationships and the contextually mediated relatedness of features. I ran three simulation studies, systematically manipulating cognitive load, encoding instructions, and the repetition and semantic content of the to-be-remembered items, to investigate the fitness and predictions of the new model. Recall of elaborated words was better than unelaborated words, which were recalled better than non-words. Recall of lists composed of items with less semantic content benefited more from repetition. The model failed to reproduce the benchmark cognitive load effect in immediate recall, but the effect returned in delayed recall, suggesting that the issue may at least in part be related to over-activation in the short-term. This dissertation lays the foundation for a body of work rich in opportunity for further development and refinement. The results add to growing evidence that the Time-Based Resource-Sharing model is underspecified and that the ancillary assumptions of a computational model are as impactful as the specification of the mechanism of interest itself.