A computational model of the two dentate gyrus blades

Abstract

The Dentate Gyrus (DG) is a key part of the hippocampus, and damage to the DG produces a wide range of pathologies, including overgeneralization of contexts, affective dysregulation (Anacker et al., 2018), and epileptogenic effects (Sloviter, 1994). The canonical model of the DG focuses on pattern separation for subsequent memory storage in the hippocampal subfield CA3. Experimental results challenge the singular focus on pattern separation and extend the function of the DG to the precise binding of objects and events to space, and the integration of information across episodes. Recent studies suggest that pattern separation and integration preferentially rely on distinct DG blades, with the suprapyramidal and infrapyramidal blades biased toward separation and integration, respectively. Here, we propose the first computational model that accounts for this distinction: an exemplar-based k-WTA architecture in the suprapyramidal DG (DGSUP) supports pattern separation and episode-specific representations, whereas an architecture with gradual heterosynaptic plasticity in the infrapyramidal DG (DGINF) supports integration of patterns across episodes. Both coding regimes are tested with two datasets: MNIST and neurally plausible entorhinal cortex inputs, thus suggesting some domain generality. Using the entorhinal cortex inputs, the two blades form place fields that either remap or maintain a stable code, consistent with experimental results. Novel inputs, including novel digit classes and novel spatial episodes, are incorporated through a neurogenesis-inspired turnover and recruitment mechanism. The two processing streams allow for a comparison of ongoing experience with the generalized expectations formed through integration across episodes. This yields prediction errors that can drive the storage of poorly predicted memories and the forgetting of well-predicted memories. The differential processing across the DG could thus aid in the iterative construction of spatial cognitive maps that encode location-dependent expectations, while at the same time preserving individual episodic memory traces. These functions are accomplished with biologically plausible learning regimes and widen the scope of DG computation beyond its well-established role in pattern separation.