Pattern separation in the hippocampus

The ability to discriminate among similar experiences is a crucial feature of episodic memory. This ability has long been hypothesized to require the hippocampus, and computational models suggest that it is dependent on pattern separation. However, empirical data for the role of the hippocampus in pattern separation have not been available until recently. This review summarizes data from electrophysiological recordings, lesion studies, immediate-early gene imaging, transgenic mouse models, as well as human functional neuroimaging, that provide convergent evidence for the involvement of particular hippocampal subfields in this key process. We discuss the impact of aging and adult neurogenesis on pattern separation, and also highlight several challenges to linking across species and approaches, and suggest future directions for investigation.

Introduction

The hippocampus is often implicated in forming new associative memories, storing memories independently of each other, retrieving memories from partial cues, and flexibly applying stored memories to novel situations. David Marr [1] was the first to suggest that recurrent collaterals (Glossary) enable a region to act as an auto-association network capable of pattern completion, the process by which incomplete or degraded representations are filled-in based on previously stored representations. Pattern completion allows for accurate generalization in the face of noise or partial sensory input. Balanced against pattern completion is the process of pattern separation, whereby similar representations are stored in a distinct, non-overlapping (orthogonalized) fashion (Figure 1a). If one were not able to perform this mnemonic discrimination, encoding new information would overwrite similar previously stored information, leading to catastrophic interference 2, 3.

Since Marr [1], emphasis has been placed on separation and completion in computational models of the hippocampus 2, 3, 4, 5, 6, 7, 8. These models suggest that the dentate gyrus (DG) granule cells are capable of performing especially strong and domain-agnostic pattern separation on the overlapping/distributed representations arriving from the entorhinal cortex (EC), projecting this signal onto the CA3 subfield of the hippocampus (Figure 1b). The CA3 receives three major excitatory inputs: (i) mossy fiber input from DG granule cells 9, 10, (ii) perforant path input directly from layer II of the EC [11], and (iii) recurrent collateral input from CA3 neurons [12] (Figure 2). The mossy fiber pathway is a powerful unidirectional input from the DG that utilizes large synapses on the proximal apical dendrites of CA3 pyramidal cells. These ‘detonator synapses’ [4] are known for their ability to strongly depolarize CA3 neurons 8, 13, 14. The extensive recurrent collateral network of CA3 has led learning theorists to postulate that the region can function as an auto-associative pattern completion network 8, 15 via attractor dynamics [16].

Neurons in layer II of the EC have collaterals that directly reach CA3, bypassing the DG [11]. This finding has led to the postulation that the mossy fiber pathway from DG to CA3 is used to force new pattern separated representations onto CA3 neurons to reduce interference and support new learning, whereas the weaker direct projection from layer II EC neurons can be used to provide a cue for recall [17]. Consistent with this idea, the inactivation of mossy fibers interferes with new learning while leaving recall intact [18]. Also, lesioning the DG input into CA3 impairs encoding but not retrieval, whereas lesioning the perforant path input directly to CA3 impairs retrieval but spares encoding [19]. Thus, it is likely that the CA3 network ‘associates’ the mossy fiber input coming from the DG granule cells with the perforant path input coming from the EC to facilitate later recall. In summary, computational models of hippocampal learning propose that the DG signal can drive activity in the CA3 along with direct EC input, with the CA3 demonstrating pattern separation signals under some circumstances and pattern completion signals under others (Figure 1b).