Dendrites help mitigate the plasticity-stability dilemma

Katharina A. Wilmes, Claudia Clopath

TL;DR

This paper proposes that the compartmentalized structure of neurons, specifically their dendrites, can help resolve the long-standing "plasticity-stability dilemma" in neural networks. By allowing plasticity to be gated independently in dendritic and somatic compartments, networks can achieve significant synaptic changes for learning while maintaining stable overall activity.

What Problem Does This Paper Address?

Neural networks learn through synaptic plasticity, where connections strengthen or weaken based on activity (Hebbian learning). However, this can lead to unstable network dynamics and the overwriting of learned information. While homeostatic mechanisms exist to stabilize activity, they are often too slow to counteract fast Hebbian plasticity. Previous solutions, like limiting plasticity or gating it globally, often sacrifice the ability to learn significantly. This paper tackles the "plasticity-stability dilemma": how can networks learn and adapt robustly without becoming unstable? They hypothesize that the complex structure of neurons, particularly the distinct dendritic and somatic compartments, offers a novel solution.

Methods

The researchers built a computational model of a balanced spiking neural network using 2-compartment pyramidal cells (soma and dendrite) and inhibitory interneurons. They implemented a standard triplet Spike-Timing-Dependent Plasticity (STDP) rule for synaptic changes and a homeostatic plasticity mechanism to maintain stable firing rates. The core of their investigation involved exploring how "gating" plasticity—modulating its strength—in different parts of the neuron (dendrite vs. soma) and through various biological mechanisms (learning rate, excitability, inhibition) impacts network stability. They quantified stability by measuring the "explosion factor" (how much firing rates increase beyond a baseline) and the critical homeostatic time constant (the slowest homeostatic response that still maintains stability). They also examined the effect of somato-dendritic coupling strength and spatially restricted gating.

Datasets

Not discussed in the paper. The study relies on a simulated neural network, not real-world datasets.

Key Insights and Ideas

  • Dendrites as independent plasticity sites: The paper's central idea is that dendrites, being electrotonically separate from the soma, can act as distinct sites for plasticity. This compartmentalization allows for targeted learning in specific neuronal regions without destabilizing the entire neuron or network.
  • Dendritic gating offers a better trade-off: Gating plasticity (e.g., increasing learning rate or excitability) in the dendrites leads to a significantly better stability-plasticity trade-off compared to gating the same mechanisms in the perisomatic (soma) region. This means more learning can occur with less risk of instability.
  • Somato-dendritic coupling is crucial: The strength of the connection between the dendrite and soma plays a critical role. Weaker coupling, mimicking states like REM sleep where memory consolidation is thought to occur, enhances the benefits of dendritic plasticity for stability.
  • Spatially precise gating improves stability: Restricting plasticity gating to only a subpopulation of neurons, rather than the entire network, further improves network stability. This suggests that localized learning signals can be more effective.

Key Results

  • Gating plasticity increases stability: Modulating learning rate, excitability, or inhibition in the network generally increases its stability, allowing for slower homeostatic mechanisms.
  • Dendritic gating is more effective: Increasing plasticity (e.g., doubling learning rate) in dendrites allowed for a significantly larger critical homeostatic time constant (meaning slower homeostatic control is sufficient) compared to the same increase in the perisomatic region. This was observed for learning rate and excitability.
  • Dendritic plasticity allows more weight change: When plasticity was gated in the dendrites, the network could accommodate larger dendritic weight changes while maintaining stability, compared to gating perisomatically.
  • Weak coupling enhances stability: Increasing somato-dendritic coupling reduced the critical homeostatic time constant, indicating that decoupling is beneficial for stability during dendritic plasticity.
  • Spatially restricted gating is advantageous: Gating plasticity in only 25% of neurons led to a much lower impact on the critical homeostatic time constant compared to gating in all cells, highlighting the stability benefits of localized learning.

Limitations and Future Work

  • Model Simplifications: The model uses simplified neuron models and plasticity rules. Future work could explore more complex neuronal architectures and plasticity mechanisms.
  • Parameter Dependence: The specific quantitative results are dependent on the chosen parameters. The authors acknowledge that the precise values for critical homeostatic time constants can vary.
  • Experimental Validation: The study is computational. Future work should aim to experimentally validate these predictions, for example, by investigating dendritic plasticity during different behavioral states or by selectively manipulating dendritic plasticity.
  • Inhibitory Plasticity: While the study touches on inhibition, further exploration of how inhibitory plasticity interacts with dendritic plasticity and somato-dendritic coupling could be valuable.
  • Memory Protection: The paper briefly mentions memory protection, suggesting that learning rate gating is most effective at preserving memories, while inhibition and excitability modulation lead to earlier memory breakdown. This could be a significant area for future research.

Summary updated

Abstract

With Hebbian learning 'who fires together wires together', well-known problems arise. Hebbian plasticity can cause unstable network dynamics and overwrite stored memories. Because the known homeostatic plasticity mechanisms tend to be too slow to combat unstable dynamics, it has been proposed that plasticity must be highly gated and synaptic strengths limited. While solving the issue of stability, gating and limiting plasticity does not solve the stability-plasticity dilemma. We propose that dendrites enable both stable network dynamics and considerable synaptic changes, as they allow the gating of plasticity in a compartment-specific manner. We investigate how gating plasticity influences network stability in plastic balanced spiking networks of neurons with dendrites. We compare how different ways to gate plasticity, namely via modulating excitability, learning rate, and inhibition increase stability. We investigate how dendritic versus perisomatic gating allows for different amounts of weight changes in stable networks. We suggest that the compartmentalisation of pyramidal cells enables dendritic synaptic changes while maintaining stability. We show that the coupling between dendrite and soma is critical for the plasticity-stability trade-off. Finally, we show that spatially restricted plasticity additionally improves stability.
BibTeX
@inproceedings{wilmes2023dendrites,
  author = {Katharina A. Wilmes and Claudia Clopath},
  title = {{Dendrites help mitigate the plasticity-stability dilemma}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-32410-0},
}
Figure