Experimental studies have found that episodic memory in humans is unreliable, often preserving little more than the gist of the experienced episode, but none of the details. We hypothesize that this property of the episodic memory system results from the fact that episodes are stored only in terms of higher order information, i.e., semantic representation, not their underlying sensory inputs. This hypothesis is similar to Tulving's SPI (Serial-Parallel-Independent) model (Tulving, 1995), in which sensory information is first processed by the semantic system before being stored in the episodic system. We have recently developed a computational model to study the interrelation between the semantic and episodic system. In this model, the semantic system compresses the high-dimensional sensory inputs to a lower dimensionality both in space and time, i.e., the semantic representation can be represented by fewer neurons and varies at a lower rate than the sensory input stream. Episodic memories are represented as sequences of semantic representations, which are stored in and retrieved from a memory network. Preliminary results indicate that episodic memory is more accurate if the underlying semantic representation has been optimized for the object that is being stored in episodic memory. The goal of this project is to extend these results to new object types and classes of objects. A good command of Python is required.
Tulving, E. (1995). Organization of memory: Quo vadis. In M. Gazzaniga (Ed.), The cognitive neurosciences (pp. 839–847). Cambridge, MA: MIT Press.