Chapter 21 Restricted Boltzmann Machine Models of Hippocampal Coding and Neurogenesis

The hippocampus has been traditionally viewed as a memorization device, creating orthogonalized representations (pattern separation) in the dentate gyrus, and performing associative retrieval (pattern completion) in the CA3. Moreover, neurogenesis in the dentate gyrus is widely assumed to increase pattern separation. Evidence that neurogenesis is important for behavioral discrimination has been erroneously taken as supporting the pattern separation assumption. Instead, we propose that the hippocampus forms a probabilistic, generative model of its input, using forward and feedback connections for encoding versus reconstruction. Using the Restricted Boltzmann Machine, we model the developmental trajectory of adult-generated neurons from hyperactive, hyperplastic, sparsely connected young neurons to less plastic, more densely connected mature neurons under tight inhibitory control. Models with neurogenesis are more robust against interference, while paradoxically generating more overlapping representations (less pattern separation). When applied to more realistic grid cell and boundary vector cell inputs, the model learns place cell representations. Finally, we simulate a full multilayer hippocampal model with neurogenesis and discuss how it can learn representations of sequential, complex events.