When shared concept cells support associations: Theory of overlapping memory engrams

Chiara Gastaldi, Tilo Schwalger, Emanuela De Falco, Rodrigo Quian Quiroga, Wulfram Gerstner

TL;DR

This paper proposes a computational model to explain how shared neurons between concept representations in the brain (hippocampus) support associations. They found that a specific range of overlap is crucial: too little prevents associations, while too much merges concepts. This work offers a theoretical framework for understanding memory recall and association chains.

What Problem Does This Paper Address?

The human brain can associate different concepts (e.g., a person and a place). Neuroscientific evidence suggests that these associations are encoded by neurons that are active for both concepts – essentially, "shared concept cells." However, it's unclear how much overlap is optimal. Too much overlap might cause the brain to treat two distinct concepts as one, hindering individual recall. Too little might not be enough to form a strong association. This paper aims to theoretically determine the optimal range of shared neurons for reliable concept association and recall, and how this mechanism might support sequential memory recall (association chains).

Methods

The authors employ an attractor neural network model to simulate memory engrams. These engrams are represented as sparse binary patterns of neuronal activity. They introduce the concept of overlapping engrams, where a certain fraction of neurons are shared between patterns.

Key methodological aspects include:

  • Mean-field theory: To analyze the network dynamics in the limit of large numbers of neurons, simplifying the complex interactions.
  • Phase-plane analysis: Used to visualize and understand the stability of different network states (e.g., single concept recall, simultaneous recall, resting state).
  • Introduction of global inhibition and adaptation: To model more complex neural dynamics, including the formation of association chains.
  • Development of algorithms for generating correlated patterns: To test how different methods of embedding overlapping memories affect network behavior and to compare with experimental data.

Datasets

Not discussed in the paper. The authors refer to experimental data from previous studies (e.g., [6]) for comparison with their model's predictions.

Key Insights and Ideas

  • The "sweet spot" for concept overlap: There's a critical range for shared neurons. Below a minimum threshold ($C_{min}$), associations are weak. Above a maximum threshold ($C_{max}$), individual concepts merge, making separate recall impossible. This highlights a crucial tradeoff in neural representation.
  • Association chains as a mechanism for free recall: The model demonstrates that with periodic modulation of inhibition (mimicking brain oscillations) and sufficient overlap, the network can transition sequentially between associated concepts, forming "association chains." This provides a theoretical basis for phenomena like free recall of word lists.
  • Context-dependent dissociation: The model predicts that introducing context (e.g., a place) can help disambiguate memories that share a common concept (e.g., a person associated with multiple places).
  • Non-hierarchical memory organization: The model's success in explaining experimental data on multi-responsive neurons suggests that memory engrams in the hippocampus might be organized in a non-hierarchical, "iterative overlap-generating" manner, rather than a strict hierarchy.

Key Results

  • Quantification of $C_{max}$: The maximal fraction of shared neurons ($C_{max}$) above which separate concept recall breaks down was determined. For typical parameters, this value was found to be around 22% (Fig 1C, bottom).
  • Minimal overlap for association: In the presence of inhibition, a minimal overlap ($C_{min}$) is required for associations to form, and this value depends on the strength of inhibition. For a specific inhibition level ($J_0 = 0.5$), $C_{min}$ was found to be around 5% (Fig 3B).
  • Association chains are possible: Simulations showed that with appropriate overlap and oscillatory inhibition, the network can successfully generate chains of recalled concepts (Fig 4B).
  • Model explains multi-responsive neuron data: The "iterative overlap-generating" algorithm, a non-hierarchical approach, best predicted the distribution of neurons responding to multiple concepts, matching experimental data (Fig 6A).

Limitations and Future Work

  • Parameter estimation: The authors note that parameters like the frequency-current curve characteristics for human MTL neurons are not precisely known, making direct quantitative prediction of $C_{max}$ challenging. They use parameters from macaque inferotemporal cortex as an approximation.
  • Simplifications in mean-field theory: While powerful, mean-field theory makes approximations, particularly regarding correlations between background patterns.
  • Computational cost of diluted networks: Simulating highly diluted networks (sparse connectivity) is computationally expensive, limiting the exploration of more biologically realistic connectivity patterns.
  • Future work: The authors suggest exploring more complex network structures, investigating the role of adaptation in more detail, and further comparing model predictions with experimental data, particularly regarding the precise mechanisms of context-dependent memory retrieval.

Summary updated

Abstract

Abstract Assemblies of neurons, called concepts cells, encode acquired concepts in human Medial Temporal Lobe. Those concept cells that are shared between two assemblies have been hypothesized to encode associations between concepts. Here we test this hypothesis in a computational model of attractor neural networks. We find that for concepts encoded in sparse neural assemblies there is a minimal fraction c min of neurons shared between assemblies below which associations cannot be reliably implemented; and a maximal fraction c max of shared neurons above which single concepts can no longer be retrieved. In the presence of a periodically modulated background signal, such as hippocampal oscillations, recall takes the form of association chains reminiscent of those postulated by theories of free recall of words. Predictions of an iterative overlap-generating model match experimental data on the number of concepts to which a neuron responds. Authors contributions All authors contributed to conception of the study and writing of the manuscript. CG and TS developed the theory. CG wrote the code for all figures. EDF and RQQ provided the experimental data. EDF and CG analyzed the data. WG and CG developed algorithms to fit the experimental data.
BibTeX
@article{gastaldi2021when,
  author = {Chiara Gastaldi and Tilo Schwalger and Emanuela De Falco and Rodrigo Quian Quiroga and Wulfram Gerstner},
  title = {{When shared concept cells support associations: Theory of overlapping memory engrams}},
  journal = {PLOS Computational Biology},
  year = {2021},
  doi = {10.1371/journal.pcbi.1009691},
}
Figure