Memory replay in balanced recurrent networks

Nikolay Chenkov, Henning Sprekeler, Richard Kempter

Memory replay—the sequential reactivation of neural patterns representing past experiences—is a hallmark of hippocampal and neocortical activity during rest. While these sequences are thought to be crucial for memory consolidation, the underlying mechanisms remain debated. Specifically, how can the relatively weak and sparse synaptic changes formed during a single experience reliably drive large-scale network activity?

The Challenge of Sparse Connectivity

Traditional models of sequence propagation, such as synfire chains, typically require dense feedforward connections to ensure robust activity transfer between groups of neurons. However, biological evidence suggests that "one-shot" learning often results in minute synaptic footprints. This paper, published in PLOS Computational Biology, proposes that the brain overcomes this limitation by leveraging pre-existing recurrent connectivity within neural assemblies to amplify weak feedforward signals.

Model Architecture and Methods

The researchers investigated this hypothesis using a combination of large-scale numerical simulations and analytical modeling. They constructed a spiking neural network (SNN) consisting of 25,000 leaky integrate-and-fire neurons (80% excitatory, 20% inhibitory). To replicate the irregular, low-rate firing observed in vivo, the network was maintained in a "balanced" state where excitation and inhibition are dynamically matched via an inhibitory plasticity rule (the Vogels et al. rule).

Schematic of the network architecture showing recurrently connected assemblies (A) linked in a sequence (B).

Within this balanced framework, the authors embedded "Hebbian assemblies"—clusters of neurons with enhanced internal recurrent connectivity ($p_{rc}$). These assemblies were then connected in a sequence through sparse feedforward projections ($p_{ff}$). By linearizing the network dynamics, they derived an "amplification factor" ($\kappa$) to predict the conditions necessary for successful sequence replay.

Key Insights: Recurrent Amplification

The study's primary finding is that recurrent connectivity within an assembly acts as a powerful signal booster. When an assembly receives a weak external cue or a sparse feedforward input from a preceding group, its internal feedback loops rapidly ramp up activity. This internal amplification allows the signal to reach the threshold required to activate the next assembly in the chain, even when the inter-assembly connections are extremely sparse.

This creates a synergistic trade-off: higher recurrent connectivity ($p_{rc}$) significantly reduces the required density of feedforward connections ($p_{ff}$). This mechanism explains how a new memory sequence can be "engraved" into the network with minimal synaptic changes, provided the individual components (assemblies) are already well-established.

Dynamics of Replay and Consolidation

The model demonstrates several biologically relevant behaviors:

  • Pattern Completion: The network is highly robust to partial cues. Stimulating as little as 20% of the first assembly in a sequence is sufficient to trigger a full, reliable replay.
  • Spontaneous Replay: If the internal amplification is strong enough, random noise fluctuations can spontaneously trigger the first assembly, leading to "hallucinated" sequences without external input.
  • Neuromodulatory Switching: By adjusting global neuronal excitability—mimicking the effects of neuromodulators like acetylcholine—the network can be toggled between a "retrieval" mode (favoring cued replay) and a "consolidation" mode (favoring spontaneous replay).
  • Bidirectional Replay: With symmetric inter-assembly connections, the model can replicate both forward and reverse replay sequences, consistent with experimental observations in rodents.

Conclusion

By integrating recurrent amplification into balanced network models, this research provides a biologically plausible framework for memory replay. It suggests that the brain's ability to consolidate memories from sparse experiences relies on the interplay between existing structural knowledge (assemblies) and new, weak associations (feedforward links). Future work will explore how these mechanisms scale when multiple, overlapping sequences are stored within the same network.

Summary updated

Abstract

Abstract Complex patterns of neural activity appear during up-states in the neocortex and sharp waves in the hippocampus, including sequences that resemble those during prior behavioral experience. The mechanisms underlying this replay are not well understood. How can small synaptic footprints engraved by experience control large-scale network activity during memory retrieval and consolidation? We hypothesize that sparse and weak synaptic connectivity between Hebbian assemblies are boosted by pre-existing recurrent connectivity within them. To investigate this idea, we connect sequences of assemblies in randomly connected spiking neuronal networks with a balance of excitation and inhibition. Simulations and analytical calculations show that recurrent connections within assemblies allow for a fast amplification of signals that indeed reduces the required number of inter-assembly connections. Replay can be evoked by small sensory-like cues or emerge spontaneously by activity fluctuations. Global—potentially neuromodulatory—alterations of neuronal excitability can switch between network states that favor retrieval and consolidation. Author Summary Synaptic plasticity is the basis for learning and memory, and many experiments indicate that memories are imprinted in synaptic connections. However, basic mechanisms of how such memories are retrieved and consolidated remain unclear. In particular, how can one-shot learning of a sequence of events achieve a sufficiently strong synaptic footprint to retrieve or replay this sequence? Using both numerical simulations of spiking neural networks and an analytic approach, we provide a biologically plausible model for understanding how minute synaptic changes in a recurrent network can nevertheless be retrieved by small cues or even manifest themselves as activity patterns that emerge spontaneously. We show how the retrieval of exceedingly small changes in the connections across assemblies is robustly facilitated by recurrent connectivity within assemblies. This interaction between recurrent amplification within an assembly and the feed-forward propagation of activity across the network establishes a basis for the retrieval of memories.
BibTeX
@article{chenkov2017memory,
  author = {Nikolay Chenkov and Henning Sprekeler and Richard Kempter},
  title = {{Memory replay in balanced recurrent networks}},
  journal = {PLOS Computational Biology},
  year = {2017},
  doi = {10.1371/journal.pcbi.1005359},
}
Figure