TL;DR
This paper introduces a large-scale spiking neural network model of the hippocampus (CA3 region) to investigate how neurogenesis and multiple plasticity mechanisms contribute to associative memory encoding and retrieval. They demonstrate that while synaptic plasticity is crucial, structural plasticity introduced by neurogenesis is essential for robust memory recall, especially when the network is heavily loaded with memories.
What Problem Does This Paper Address?
The hippocampus, particularly the CA3 region, is known to be critical for forming and retrieving long-term declarative memories. These memories are thought to be stored as "engrams" within the highly recurrent CA3 network. However, existing spiking neural network models of CA3 have been too small-scale and lacked biological realism to adequately study memory encoding and retrieval. Furthermore, the role of neurogenesis (the birth of new neurons) in the dentate gyrus and its impact on CA3 memory processing remained largely unexplored in these realistic models. This paper aims to bridge this gap by developing a biologically plausible, large-scale spiking network model that incorporates both synaptic plasticity and neurogenesis to understand their combined effects on memory.
Methods
The authors developed a three-layer spiking neural network model comprising the entorhinal cortex (EC), dentate gyrus (DG), and CA3 regions. They utilized adaptive integrate-and-fire (AIF) neurons to simulate the network.
Key methodological aspects include:
- Large-Scale Network: The CA3 region was modeled with 3200 excitatory and 800 inhibitory neurons, significantly larger than previous models.
- Multiple Plasticity Models: They investigated three types of excitatory plasticity (pair STDP, triplet STDP, and a unified plasticity model) and inhibitory plasticity. The unified plasticity model, incorporating Hebbian, heterosynaptic, and transmitter-induced mechanisms, proved most effective.
- Neurogenesis Simulation: Neurogenesis was simulated by making newly born DG neurons more excitable and by recycling DG neurons to encode new memories, mimicking "turnover homeostasis." This contrasts with a "non-neurogenesis" setting where DG neurons retain their properties and connections.
- Encoding and Retrieval Phases: Memories were encoded by stimulating EC neurons, which triggered activity cascades through the DG to CA3. Retrieval involved reactivating a subset of EC neurons, with only the EC-CA3 pathway active.
- Signal-to-Noise Ratio (SNR): A quantitative measure, SNR, was used to assess the quality of memory retrieval by comparing the set of activated CA3 neurons during retrieval to those activated during encoding.
Datasets
Not discussed in the paper. The study uses simulated data generated by the spiking neural network model.
Key Insights and Ideas
- Synaptic Plasticity Alone is Insufficient: While various synaptic plasticity rules are necessary for encoding and retrieval, they are not enough on their own, especially when the network is heavily loaded.
- Neurogenesis is Crucial for Robust Recall: The introduction of neurogenesis (simulated as increased excitability of new DG neurons and recycling of DG neurons) significantly enhances memory retrieval, particularly for recent memories, and prevents the complete loss of older ones.
- "Forgetting" Mechanisms Differ: In the neurogenesis case, forgetting is attributed to the deletion of connections. In the absence of neurogenesis, forgetting is driven by increased inhibition that effectively erases memory engrams.
- Recurrent Connections Aid Pattern Completion but Amplify Noise: While recurrent (CA3-CA3) connections improve retrieval, especially for partial cues, they also amplify network noise, which can degrade performance when the network is highly stimulated or noisy.
- Inhibitory Plasticity Balances Excitation: The model highlights the critical role of inhibitory plasticity in maintaining network stability, especially in the absence of neurogenesis, where it must counteract increased excitation from recruiting more CA3 neurons for each new memory.
Key Results
- The unified plasticity model demonstrated superior performance in memory encoding and retrieval compared to pair and triplet STDP.
- With neurogenesis, the network could successfully encode and retrieve up to 200 memories, with better retrieval of more recent memories.
- Without neurogenesis, memory retrieval deteriorated significantly with an increasing number of encoded memories, with the SNR dropping to levels comparable to random chance for older memories.
- For 32 EC neurons per stimulus, recurrent plasticity (FF+REC) provided better SNR than feedforward plasticity (FF) alone, indicating pattern completion benefits.
- However, with more EC neurons per stimulus (96 and 160), recurrent plasticity amplified noise, leading to poorer SNR compared to FF plasticity alone.
- The study systematically analyzed the impact of different noise sources (input noise, DG noise, EC noise, CA3 noise) on SNR, finding that while DG noise had minimal impact, EC and CA3 noise significantly affected retrieval quality.
Limitations and Future Work
- Biological Realism: While the model is more biologically plausible than previous ones, it is still a simplification. For instance, the simulation of neurogenesis through increased excitability and recycling is a functional approximation.
- Forgetting Mechanism: The paper notes that forgetting in the neurogenesis case is due to connection deletion, which is a simplified representation of biological processes.
- Cortical Consolidation: The study focuses solely on the hippocampus and does not model the process of memory consolidation in the cortex, which is a crucial aspect of long-term memory.
- Future Work: The authors suggest that the model provides a framework for further extension in computational studies of the hippocampus's role in memory. They also highlight the need for further research into how the hippocampus functions as part of a complete system with the cortex. The role of recurrent connections in pattern completion and noise amplification warrants further investigation.
Summary updated