Prediction and memory: A predictive coding account

Helen C. Barron, Ryszard Auksztulewicz, Karl FristonProgress 2020 · DOI

TL;DR

This paper proposes a unified predictive coding framework for understanding the hippocampus's dual roles in memory recall and online prediction. It suggests that the hippocampus orchestrates distinct interactions with the neocortex – inhibitory for prediction and facilitatory for recall – by modulating the "precision" of prediction error signals, mediated by neuromodulators and inhibitory interneurons.

What Problem Does This Paper Address?

The hippocampus is known for its critical role in forming and retrieving rich, detailed episodic memories. Simultaneously, it's implicated in predicting upcoming sensory experiences. However, within the popular predictive coding framework, these two functions appear to have opposing effects on neocortical processing: memory recall seems to involve excitation to reinstate past experiences, while prediction involves inhibition to suppress prediction errors. This paper aims to resolve this apparent contradiction by proposing a unified account.

Methods

This is a hypothesis-driven theoretical paper. The authors leverage existing literature from neuroscience, computational modeling, and predictive coding theory. They propose a conceptual model that integrates findings on hippocampal anatomy, neuronal circuitry (particularly inhibitory interneurons), neuromodulation (like acetylcholine), and the principles of predictive coding. They don't present new experimental data but rather synthesize existing evidence to build their theoretical framework and propose testable predictions.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Unified Predictive Coding Framework: The core idea is that both memory recall and prediction are governed by the same underlying predictive coding principles, but manifest through different neural mechanisms.
  • Inhibition vs. Disinhibition: The paper proposes that the hippocampus interacts with the neocortex via two distinct pathways:
  • Prediction: Descending predictions from the hippocampus inhibit neocortical prediction error signals, effectively "explaining away" expected sensory input. This is mediated by direct inhibition of pyramidal cells or through inhibitory interneurons.
  • Recall: Memory recall involves disinhibition. The hippocampus facilitates the reinstatement of neocortical activity patterns by disinhibiting specific neuronal populations, effectively increasing their gain. This is proposed to be mediated by VIP+ interneurons.
  • Precision as the Key Modulator: The authors argue that the "precision" of prediction error signals (how much weight they are given) is the crucial factor determining whether hippocampal output is inhibitory or facilitatory. Higher precision amplifies prediction errors, leading to recall, while lower precision suppresses them, facilitating prediction.
  • Neuromodulation's Role: Neuromodulators like acetylcholine are proposed to control this precision, influencing the balance between inhibition and disinhibition in the neocortex.
  • Memory Recall as "Fictive" Prediction Errors: In the absence of sensory input, memory recall is framed as generating "fictive" prediction errors. These offline errors act as training signals to update the brain's generative model, optimizing it for future predictions, akin to machine learning algorithms.

Key Results

As a theoretical paper, there are no quantitative experimental results. The "results" are the conceptual framework and the proposed mechanisms:

  • A theoretical model explaining how the hippocampus can both predict sensory input (via inhibition) and recall memories (via disinhibition).
  • Identification of specific neuronal populations (VIP+, PV+, SOM+ interneurons) and neuromodulatory systems (acetylcholine) as key players in mediating these distinct hippocampal-neocortical interactions.
  • A set of testable predictions (detailed in Box 2) that can guide future experimental research.

Limitations and Future Work

The authors acknowledge several points:

  • Empirical Verification: While the model integrates existing findings, direct experimental evidence for all proposed mechanisms (e.g., specific roles of interneuron subtypes in hippocampal-neocortical interactions during prediction vs. recall) is still needed.
  • Cellular Specificity: It remains to be definitively established whether the same hippocampal cells are involved in both prediction and memory recall.
  • Neuromodulatory Complexity: The precise interplay of various neuromodulators and their specific roles in setting precision and gating hippocampal output requires further investigation.

Future work should focus on:

  • Using optogenetic and chemogenetic tools to selectively manipulate hippocampal projections and specific interneuron populations in the neocortex to test the proposed inhibitory and disinhibitory pathways.
  • Investigating the role of neuromodulators, particularly acetylcholine, in modulating hippocampal-neocortical interactions during prediction and recall tasks.
  • Employing advanced neuroimaging techniques (e.g., representational fMRI) to examine how offline memory reinstatement (fictive prediction errors) might update higher-level cortical representations.
  • Exploring the implications of this framework for understanding disorders of memory and prediction.

Summary updated

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.
BibTeX
@inproceedings{barron2020prediction,
  author = {Helen C. Barron and Ryszard Auksztulewicz and Karl Friston},
  title = {{Prediction and memory: A predictive coding account}},
  booktitle = {Progress},
  year = {2020},
  doi = {10.1016/j.pneurobio.2020.101821},
}

Loading related content...

Figure