TL;DR
This paper introduces a novel voltage-based inhibitory synaptic plasticity (vISP) model that regulates neural network activity by targeting a specific membrane voltage. Unlike previous models that often suppress neuronal diversity, vISP allows for a broad range of firing rates and heterogeneous neuronal dynamics, better reflecting biological observations in brain regions like the hippocampus.
What Problem Does This Paper Address?
Biological neural networks exhibit remarkable diversity in neuronal activity and structure, which is thought to be crucial for complex functions like information processing and memory. Homeostatic mechanisms, such as inhibitory synaptic plasticity, are known to maintain network stability. However, existing computational models of inhibitory plasticity often impose strict constraints, leading to a loss of neuronal diversity and an inability to reproduce the wide range of firing rates and functional specializations observed in real brains. This paper aims to bridge this gap by proposing a new inhibitory plasticity mechanism that can maintain network homeostasis while preserving and even promoting neuronal diversity.
Methods
The core of this work is the development of a voltage-dependent inhibitory synaptic plasticity (vISP) model. In this model, the update of inhibitory synaptic weights is not solely dependent on spike timing (as in traditional spike-based models) but also on the postsynaptic neuron's membrane voltage. Specifically, the vISP rule aims to maintain the average postsynaptic membrane voltage at a target value over a long timescale. This is achieved by incorporating a presynaptic trace and a low-pass-filtered version of the postsynaptic membrane potential.
The authors then tested this vISP model in various network architectures:
- Feedforward networks: To assess how vISP regulates postsynaptic activity and the excitatory/inhibitory (E/I) ratio. They compared it directly with a spike-based inhibitory plasticity (sISP) model.
- Feedforward networks with correlated inputs: To investigate the development of co-tuned excitatory and inhibitory receptive fields, mimicking observations in auditory cortex.
- Recurrent networks: To explore its role in memory formation and retrieval and its impact on network heterogeneity.
- Hippocampal CA1 network model: To account for the diverse place cell properties observed experimentally, incorporating heterogeneous neuronal excitability.
Datasets
Not discussed in the paper. The study relies on simulated neural networks and their parameters.
Key Insights and Ideas
- Preserving Diversity: The most significant insight is that vISP can regulate network activity and maintain homeostasis without enforcing a narrow target firing rate, unlike many previous models. This allows for a broad distribution of firing rates, crucial for biological realism.
- Flexible Homeostasis: vISP sets a target for the E/I ratio and a maximum firing rate, but not a strict minimum. This flexibility allows neurons to be silent or exhibit varying levels of activity, contributing to diverse coding strategies (sparse and dense).
- Adaptation to Change: The vISP model demonstrates an ability to adapt inhibitory connections to changes in excitatory inputs, allowing the network to re-establish balance and co-tuning even after significant shifts in excitatory receptive fields.
- Memory Dynamics: In recurrent networks, vISP supports memory formation and retrieval by allowing the "memory assembly" to maintain a distinct, higher firing rate compared to the rest of the network, rather than being suppressed back to baseline activity as seen with sISP.
- Explaining Place Cell Variability: When combined with heterogeneous neuronal excitability, vISP successfully reproduces the wide range of place cell properties (number of fields, field size, coverage) observed in the hippocampus.
Key Results
- In feedforward networks, vISP regulates the E/I ratio and sets a maximum postsynaptic firing rate, but allows firing rates to span a wide range below this maximum. In contrast, sISP imposes a narrow target firing rate.
- With correlated excitatory and inhibitory inputs, vISP leads to the development of co-tuned excitatory and inhibitory receptive fields, similar to experimental findings.
- In recurrent networks, vISP supports memory formation by allowing a memory assembly to maintain a differentiated activity level, and this activity level is broadly distributed among neurons within the assembly.
- In a hippocampal CA1 model with heterogeneous excitability, vISP reproduces the observed diversity in place cell features, including firing rates, number of fields, field sizes, and coverage.
Limitations and Future Work
- Experimental Validation: While the model reproduces many experimental observations, further experimental validation of the specific voltage-dependence of inhibitory plasticity is needed.
- Neuron Model Simplification: The study uses simplified neuron models (AdEx and LIF). More biologically realistic models with detailed dendritic integration could reveal further nuances.
- Targeting Site: The paper suggests that the targeting site of inhibitory synapses (e.g., soma vs. dendrites) might influence the effects of vISP, a hypothesis that warrants further investigation with more detailed neuron models.
- Combined Plasticity Rules: The authors mention that their vISP model could be combined with other plasticity rules, such as input-dependent inhibitory plasticity (IDIP), to achieve even more sophisticated network control.
- Context-Dependent Dynamics: The flexibility of vISP suggests potential for encoding context-dependent information, which could be explored further.
Summary updated