The Dentate Gyrus (DG) is a central hippocampal structure implicated in multiple dysfunctions when damaged, including context overgeneralization, affective dysregulation, and epileptogenic effects. While canonical accounts emphasize pattern separation in the DG to prepare inputs for CA3 memory storage, accumulating experimental evidence suggests a broader set of functions, such as precise binding of objects and events to spatial locations and integration of information across episodes. This work presents a computational model that formalizes a functional division across the two DG blades to account for these observations.
Recent empirical studies propose that the two anatomical blades of the DG are biased toward different computations: the suprapyramidal blade favors pattern separation, producing episode-specific representations, whereas the infrapyramidal blade favors integration across episodes, producing generalized codes. The model described here is explicitly designed to capture this blade-specific biasing and to test its computational consequences for memory formation and spatial coding.
The proposed model implements two parallel processing streams corresponding to the suprapyramidal DG (DGSUP) and the infrapyramidal DG (DGINF). DGSUP uses an exemplar-based k-Winner-Take-All (k-WTA) architecture to promote sparse, discriminative representations suitable for separation of similar inputs. DGINF uses a learning regime with gradual heterosynaptic plasticity to encourage integration of patterns encountered across multiple episodes. Both streams are fed the same inputs, and their outputs can be compared to generate prediction errors that influence memory storage.
In the model, DGSUP implements an exemplar-based k-WTA mechanism. This architecture selects a small subset of active units (winners) in response to an input, promoting sparse and distinct activation patterns across similar inputs. The exemplar-based aspect supports the formation of episode-specific representations, aligning with the blade’s bias toward pattern separation and the preservation of individual episodic traces.
DGINF is instantiated with a slower, heterosynaptic plasticity rule that adjusts synaptic strengths gradually across repeated experiences. This architecture supports the emergence of integrated, generalized representations that combine information across episodes. The gradual nature of plasticity in DGINF contrasts with the fast, exemplar-driven encoding in DGSUP, allowing the model to represent both stable expectations and accumulated statistics of past inputs.
The two coding regimes were tested on two distinct datasets: the MNIST digit dataset and neurally plausible inputs derived from the entorhinal cortex. Use of these diverse inputs was intended to probe domain generality of the proposed mechanisms. Both DGSUP and DGINF produced meaningful representations in each domain, supporting the idea that the blade-specific architectures are not limited to a single input modality.
When driven by entorhinal cortex inputs, the two blades developed place-field-like responses. Importantly, these place fields exhibited differential behavior: one blade produced place fields that remapped across episodes while the other maintained a more stable code. This behavior in the model is consistent with experimental observations linking blade-specific activity patterns to different remapping tendencies.
To handle novel inputs—such as new digit classes in MNIST or novel spatial episodes—the model implements a neurogenesis-inspired turnover and recruitment mechanism. This process allows replacement or recruitment of representational units so that novel stimuli can be incorporated into the network without catastrophic interference of existing representations.
The dual-stream architecture permits direct comparison between ongoing, episode-specific representations generated by DGSUP and the generalized expectations formed by DGINF. Discrepancies between the two streams produce prediction errors. The model uses these errors to bias memory decisions: poorly predicted (high prediction error) experiences are preferentially stored as new episodic traces, while well-predicted (low prediction error) experiences are more likely to be forgotten. This mechanism links blade-specific computation to selective memory retention and forgetting.
The authors emphasize that the learning regimes used are biologically plausible, and that the model broadens the conceptual scope of DG computation beyond canonical pattern separation. By combining episode-specific pattern separation with across-episode integration, the two-blade architecture supports iterative construction of spatial cognitive maps that encode location-dependent expectations while preserving individual episodic memory traces. These functions may help explain experimental findings on blade-specific place-field dynamics and suggest a mechanistic role for DG heterogeneity in memory and spatial cognition.
Funding sources reported include the Max Planck Society, Max Planck School of Cognition, the European Research Council, The Kavli Foundation, and the Kristian Gerhard Jebsen Foundation Helse Midt Norge. The authors declared no competing interests.