Can We Build an Artificial Hippocampus?

Artem Kirsanov
Apr 30, 2023
4 notes
4 Notes in this Video

Factorized World Models and the Tolman-Eichenbaum Machine Prediction Objective

WorldModels Generalization PredictionObjective
03:30

The Tolman-Eichenbaum Machine (TEM) is a biologically inspired model that learns internal world models by predicting observations from sequences of actions and sensory inputs, using hippocampal organization as a blueprint.

Position Module and Path Integration as Grid Cell Analog

PathIntegration GridCells PositionModule
10:00

TEM’s position module serves as a computational analog of medial entorhinal cortex, updating an internal location code based solely on action sequences, much like biological grid cells integrate self-motion.

Memory Module as Place Cell Analog and Remapping Mechanism

PlaceCells Remapping ConjunctiveCoding
15:30

TEM’s memory module binds positional codes with sensory inputs, functioning as an associative store analogous to hippocampal place cells that fire in specific locations and remap between environments.

Statistics of Experience and Salient Hotspots in Learned Maps

ExperienceStatistics SalienceHotspots RepresentationBias
20:30

TEM’s learned representations reflect not only structural constraints of environments but also the statistics of experience, paralleling how animals over-sample boundaries, safe zones, and reward locations.