Prediction and memory: A predictive coding account

Abstract

The hippocampus is crucial for episodic memory, but it is also involved in online prediction. Evidence suggests that a unitary hippocampal code underlies both episodic memory and predictive processing, yet within a predictive coding framework the hippocampal-neocortical interactions that accompany these two phenomena are distinct and opposing. Namely, during episodic recall, the hippocampus is thought to exert an excitatory influence on the neocortex, to reinstate activity patterns across cortical circuits. This contrasts with empirical and theoretical work on predictive processing, where descending predictions suppress prediction errors to 'explain away' ascending inputs via cortical inhibition. In this hypothesis piece, we attempt to dissolve this previously overlooked dialectic. We consider how the hippocampus may facilitate both prediction and memory, respectively, by inhibiting neocortical prediction errors or increasing their gain. We propose that these distinct processing modes depend upon the neuromodulatory gain (or precision) ascribed to prediction error units. Within this framework, memory recall is cast as arising from fictive prediction errors that furnish training signals to optimise generative models of the world, in the absence of sensory data.

Keywords: Hippocampus; Interneuron; Memory; Neocortex; Prediction.

Fig. 1

The hippocampus as both a memory index and a generative model. A) Schematic illustrating the hippocampus as a memory index: During memory recall, activity patterns across neocortex are reinstated to recapitulate previous sensory experience (shown in red, distributed across the neocortical hierarchy). The hippocampus (shown in blue), which is anatomically situated at the top of a cortical processing hierarchy, is thought to orchestrate this reinstatement by binding and linking activity patterns stored across distributed neocortical networks. B) When rodents repeatedly navigate on a one-dimensional track (shown in grey), spatially tuned principal cells in the hippocampus (shown in red) show a backward skew in their firing rate (filled line) relative to the first run on the linear track (dotted line) (schematic adapted from Mehta et al., 1997). This backward skew can be explained by a Successor Representation (Stachenfeld et al., 2017) where the hippocampus represents upcoming locations or states that are reliably predicted from the current location or state. C–D) Schematic showing neocortex at an intermediary level in the cortical hierarchy. Within a predictive coding framework, the dual aspect role of the hippocampus gives rise to two complementary hippocampal-neocortical interactions. Descending inputs from the hippocampus are shown in blue. An example subset of cells in the neocortex are shown in the black box with low firing rate indicated in pale pink and high firing rate indicated in red. Ascending sensory input (or prediction error signals) are shown in green. C) As a generator of predictions, or generative model, the hippocampus accumulates ascending prediction errors from neocortical neurons lower in the hierarchy (not shown) and responds with descending predictions to neocortex that inhibit the neocortical prediction error signals. Left-hand panel: When the sensory input is unexpected, the resulting prediction errors are represented in the neocortical hierarchy. Right-hand panel: With learning, the hippocampal generative model is updated until the hippocampal predictions ‘explain away’ prediction errors by suppressing neocortical activity. D) As a memory index, the hippocampus provides descending input to the neocortex to selectively reinstate activity patterns that recapitulate previous sensory experience. The hippocampal memory index can thus facilitate neocortical activity, even in the absence of sensory input.

Fig. 2

Inhibition and disinhibition within the canonical neocortical circuit motif. A) Schematic showing a circuit motif that employs disinhibition (i.e. inhibition of inhibition). Typically, VIP+ interneurons provide disinhibitory control by targeting PV+ and/or SOM+ interneurons that otherwise inhibit the target excitatory principal neurons. VIP, PV and SOM refer to VIP+, PV+ and SOM+ respectively. Interneurons are shown in grey. Pyramidal cells are shown in red. B–E) Disinhibition of human neocortex leads to re-expression of associative memories formed between visual stimuli that are rotationally invariant. Furthermore, for two overlapping memories, disinhibition in human neocortex increases memory interference. Adapted from Koolschijn et al., 2019. B) Schematic showing how transient disinhibition of human neocortex can be achieved using unilateral anodal transcranial direct current stimulation (tDCS), with the anodal electrode positioned above a target region, the anterior lateral occipital cortex (LOC), which has previously been shown to encode associations between visual stimuli that are rotationally invariant (Barron et al., 2016). The cathodal electrode was positioned over the contralateral supraorbital ridge. C) When tDCS is applied for 20 minutes using the configuration shown in B, a reduction in the concentration of neocortical GABA is observed in anterior LOC, measured with Magnetic Resonance Spectroscopy (MRS). D) Left: average position of the anodal tDCS electrode, projected into the brain (red-yellow, with group average in yellow) and average position of the MRS voxel (blue) from which the change in concentration of GABA was measured. Right: when neocortical GABA is reduced using brain stimulation, functional Magnetic Resonance Imaging (fMRI) reveals re-expression of associative memories and an increase in memory interference in the brain region underneath the anodal electrode. E) Underneath the anodal electrode, an increase in associative memory expression measured with fMRI can be observed during application of tDCS (providing effective disinhibition), suggesting that expression of associative memories is otherwise quenched by cortical inhibition. F) Prediction signalling in different domains affects the gain in sensory cortical regions, expressed as interactions between ensembles of superficial pyramidal cells (SP) and inhibitory interneurons (IN). However, the exact neuromodulatory mechanisms are domain-specific: ‘what’-predictions are mediated by NMDAR-dependent short-term plasticity contingent on the postsynaptic effects of descending connections from deep pyramidal cells (DP) of higher-order regions, such as the hippocampus, on SP of lower-order regions; ‘when’-predictions are instead subserved by classical (e.g., dopaminergic, DA, or cholinergic, ACh) modulation of postsynaptic gain in lower-order sensory regions. When interacting together, these temporal predictions could be specific to a particular stimulus content (Auksztulewicz et al., 2018).

Fig. 3

Schematic showing the proposed neuronal architecture underlying inhibitory and facilitatory hippocampal-neocortical interactions. Within the neocortical hierarchy, message passing is orchestrated by a canonical microcircuit that includes both excitatory (red and black) and inhibitory (beige) cells. In the superficial layers of each cortical level, superficial pyramidal cells (red) compare the activity of representational units (black) with top-down predictions relayed via SOM+ inhibitory interneurons (SOM). These interneurons are targeted by descending prediction signals that originate in deep pyramidal cells (black) from the level above. The mismatch between representations and descending predictions (black lines) constitutes a prediction error. This prediction error signal (red lines) is passed back up the cortical hierarchy and is received by prediction units (black) that drive responses in higher representational units, or, at the apex of the processing hierarchy, in the hippocampus. Therefore, as information moves up the cortical processing hierarchy, sensory input is replaced by prediction error signals that convey the only information yet to be explained. These prediction error signals drive representations in higher levels of the cortical hierarchy to provide better predictions, but also drive associative plasticity to update internally generated predictions that in the hippocampus draw on memory. The output from the hippocampus targets neocortex via glutamatergic projections to deep pyramidal cells (black, e.g. in the entorhinal cortex), or via long-range GABAergic projections to superficial cells (e.g. in retrosplenial cortex Yamawaki et al., 2019a; not shown here). Using a predictive coding framework, we propose that the hippocampus uses a unitary code with a dual aspect function. This dual aspect function can be characterised as follows: During prediction, the hippocampus can provide multi-sensory predictions to ‘explain away’ prediction errors at lower levels of the cortical hierarchy. This manifests as an inhibitory hippocampal-neocortical interaction – here mediated by SOM+ inhibitory interneurons. During memory recall, the hippocampus can provide a memory index to neocortex, to selectively reinstate activity patterns across distributed neocortical networks, which manifests as a facilitatory hippocampal-neocortical interaction – here mediated polysynaptically via VIP+ and SOM+/PV+ inhibitory interneurons. We propose that the diversity of inhibitory interneurons – and their selective responses to classical neuromodulators or NMDAR-mediated stimulation– provide the necessary machinery for complementary inhibitory and facilitatory hippocampal-neocortical interactions. Computationally, the facilitatory (disinhibitory) effect of hippocampal projections would, in this scheme, encode the precision of prediction error units by modulating their postsynaptic excitability. For simplicity, we have omitted many connections and cell types in the canonical microcircuit (e.g., spiny stellate cells in layer 4) and in the hippocampus. Furthermore, we have omitted descending projections directly to PV+ interneurons. Excitatory synapses are denoted with lines ending in a circle, while inhibitory synapses are denoted by a diamond. Note that superficial pyramidal cells receive excitatory and inhibitory influences that underwrite a prediction error, while the precision of the encoded prediction error is controlled by modulatory (orange) interactions with VIP+ inhibitory interneurons. ACh refers to acetylcholine. PV, SOM and VIP refer to PV+, SOM+ and VIP+ interneurons. DG refers to the dentate gyrus, Sub refers to subiculum, which together with CA1 and CA3 constitute subfields of the hippocampus that reside along the performant pathway; ‘n’ refers to the level in the cortical hierarchy.