Based on empirical results, we are developing a model of how humans form episodic associations between words, and how such associations are affected by existing semantic memory. A spreading activation process on a self-organizing map of word meanings matches the experiments well, and leads to several interesting predictions about neighborhood effects and impaired association performance under schizophrenia.
Inspired by Damasio's convergence-zone idea, the inputs to the memory are assumed to be represented locally in perceptual maps, and the memory encoding is a sparse random pattern in the hippocampus. Such a memory can be analyzed mathematically and simulated computationally, and it suggests how the hippocampal memory can have a high capacity even with sparse connectivity and a relatively small number of computational units. One-shot storage is shown to require large learning rates, and temporary storage (during transfer to neocortex) possible through weight normalization.
The Trace Feature Map is a self-organizing map where lateral connections between units are used to encode a memory trace: the map remembers that at some point, an input was received that was mapped at a particular location on the map. The Trace Map model was originally developed as the episodic memory component for the DISCERN story processing system, and it exhibits plausible cognitive behavior for this task. More recent and more unique traces are easier to recall, and the memory capacity degrades gradually when overloaded.