TL;DR
This paper introduces a computational model that explains how the brain links memories of events that occur close in time. By incorporating neural excitability alongside traditional synaptic plasticity, the model demonstrates that a neuron's inherent "excitable" state can bias it towards storing new memories, leading to overlapping neural representations for temporally proximate events. This work sheds light on the mechanisms underlying memory consolidation and linking.
What Problem Does This Paper Address?
The brain's ability to form and retain memories is remarkable, but the precise neural mechanisms underlying this process, particularly how memories are stored and linked, are still being unraveled. While synaptic plasticity (changes in the strength of connections between neurons) has been a primary focus, recent experimental evidence suggests that intrinsic neural excitability (a neuron's inherent tendency to fire) also plays a crucial role. Specifically, studies have observed that memory engrams (groups of neurons representing a memory) for events that happen close in time tend to overlap, facilitating memory linking. However, the exact mechanisms by which intrinsic excitability contributes to this overlap and memory allocation remain unclear. This paper aims to bridge this gap by developing a computational model that integrates both synaptic plasticity and neural excitability to explain these phenomena.
Methods
The researchers developed a rate-based recurrent neural network (RNN) model. Key features of their approach include:
- Recurrent Neural Network Architecture: A standard RNN structure with feed-forward inputs and all-to-all recurrent connections.
- Synaptic Plasticity: Implemented using a Hebbian learning rule, where synaptic weights strengthen if pre- and post-synaptic neurons are active simultaneously.
- Intrinsic Neural Excitability: This is modeled as a time-varying threshold for each neuron. When a neuron's firing rate exceeds an "active" threshold, its excitability increases temporarily before decaying back to its baseline. This dynamic excitability is hypothesized to be regulated by factors like the CREB transcription factor.
- Feed-Forward Inputs: Represent different "contexts" or events, each activating a specific subpopulation of neurons.
- Inhibition: A global inhibition mechanism is included to model competition between neurons for memory allocation.
The novelty lies in the explicit integration of dynamic, activity-dependent intrinsic neural excitability as a complementary mechanism to synaptic plasticity for memory engram formation and overlap.
Datasets
Not discussed in the paper. The study uses a simulated environment and does not rely on external datasets for training or evaluation.
Key Insights and Ideas
- Excitability as a Memory Allocator: The model demonstrates that a neuron's intrinsic excitability can act as a bias, making it more likely to be recruited into a new memory engram. Neurons that are already more excitable are preferentially selected.
- Time-Dependent Overlap: When two events (contexts) are presented close in time, the model shows structural and functional overlap between their corresponding memory engrams. This overlap is significantly reduced when the temporal delay between events increases.
- Linking Memories: The structural overlap between engrams of temporally close events naturally leads to memory linking. Recalling one memory can activate parts of the other, transferring associated information (e.g., fear responses in their simulation).
- Inhibition as a Competitor Regulator: The level of inhibition in the network plays a crucial role in regulating competition among neurons. Higher inhibition reduces engram overlap, suggesting a delicate balance is needed for memory linking.
- Testable Predictions: The model generates specific, testable predictions about how manipulating excitability and inhibition should affect memory formation and linking, which can be experimentally verified.
Key Results
- The model successfully reproduces the formation of memory engrams and their reactivation during recall, consistent with experimental observations.
- When two contexts are presented with a short temporal delay (e.g., 6 hours), the model generates overlapping memory engrams. This overlap is significantly less pronounced when the delay is longer (e.g., 24 hours).
- In a fear conditioning simulation, the model showed that temporally linked memories (separated by 5 hours) exhibit memory linking (fear response transfer), while memories with a longer delay (7 days) do not.
- Artificially enhancing the excitability of a subset of neurons before training biased memory allocation towards those neurons, and inhibiting them during recall reduced memory strength.
- Inhibiting neurons that received enhanced excitability during the formation of the first memory also reduced the overlap between the first and second memory engrams when they were temporally close.
Limitations and Future Work
- Simplified Excitability Dynamics: The model uses a simplified representation of excitability dynamics, primarily driven by a transcription factor like CREB. Future work could incorporate other known regulators of neural excitability.
- Homogeneous Inhibition: The model uses a homogeneous global inhibition mechanism. Exploring the effects of different types of inhibitory neurons and their specific roles could provide a more nuanced understanding.
- No Explicit Pattern Separation: While the model can form memories, it did not explicitly evaluate its ability to perform pattern separation (distinguishing similar but distinct memories). This is suggested as an interesting avenue for future research.
- Static Recurrent Weights: The model's recurrent weights are static when the network is not stimulated, meaning the delay between training and recall doesn't impact weights in the absence of stimulation. Investigating how engrams evolve over longer timescales is also suggested.
- Conceptual Relation Ignored: The current model's memory linking is driven solely by temporal proximity, not by the conceptual relationship between events. Future work could explore how conceptual similarity influences engram overlap and linking.
Summary updated