| In this dissertation we develop a new neural network model that explains how the hippocampus can implement the functions of episodic memory and novelty detection for spatiotemporal patterns. We review relevant literature from psychology, neuroscience, and computer science that forms the foundations of our proposal. We also review existing models of novelty detection and hippocampal function and present several open questions that we answer in the dissertation. The first model that we develop and analyze explains how a resonance network modeled on hippocampal subregion CA3 can implement the function of novelty detection. The second model extends the first to include the function of episodic memory within a more biologically motivated architecture. We conduct a series of experimental simulations that test this model on issues such as long-range pattern completion, variable transmission delays, pattern stutter, novelty-contingent encoding, partial conditioning, subsequence matching, and blocking. We find that increasing the variability of the transmission delays and increasing the pattern stutter improve recall performance. We find that novelty-contingent encoding is beneficial whenever multiple patterns are stored, but especially when they appear with different frequencies. We demonstrate how novelty-contingent encoding produces blocking in the network. We show that during a subsequence matching task, the network can both complete the full pattern and indicate that the recalled pattern was stored as two separate patterns. We conclude with a discussion on broader issues relevant to the model. |
