Introduction: The Stability-Plasticity Paradox
A fundamental challenge in neuroscience is understanding how the brain learns and stores memories without succumbing to "runaway" neural activity. While Hebbian learning—often summarized as "cells that fire together, wire together"—is a cornerstone of neural plasticity, it is inherently unstable in computational models. Strengthening synapses increases firing rates, which in turn further strengthens synapses, creating a positive feedback loop that can saturate the entire network.
Furthermore, biological systems must navigate the "stability-plasticity dilemma": the ability to learn new information without overwriting existing memory traces. Standard Spike-Timing-Dependent Plasticity (STDP) often fails here, as new patterns can interfere with and erase established memories.
Orchestrating Plasticity Across Timescales
To address these issues, this research proposes "orchestrated plasticity"—a combination of Hebbian and non-Hebbian rules acting across a hierarchy of timescales. While homeostatic mechanisms like synaptic scaling are well-known, they typically operate over hours or days. This model highlights the necessity of fast non-Hebbian mechanisms acting on the scale of seconds to provide immediate stabilization.
Methods: The Spiking Network Model
The authors simulated a network of 5,120 spiking neurons (4,096 excitatory and 1,024 inhibitory) using the Auryn simulator. The core of the model is a unified weight-update equation that integrates several key mechanisms:
- Triplet STDP: A Hebbian rule that considers three-spike interactions, providing a more biologically accurate representation of plasticity than standard pair-based STDP.
- Non-Hebbian Stabilizers: Heterosynaptic and transmitter-induced terms act as "brakes" to prevent runaway excitation.
- Synaptic Consolidation: A "slow" reference weight tracks the "fast" synaptic weight, acting as a long-term anchor that protects memories from transient interference.
The network was trained using structured spatial stimuli (e.g., geometric shapes) presented as Poisson spike trains to develop receptive fields and memory assemblies.
Key Results: Stable Memory and Associative Recall
The orchestrated plasticity rule successfully enabled the formation of stable neural assemblies. Key findings include:
- Stable Selective Delay Activity: After a stimulus is removed, the corresponding assembly maintains an elevated firing rate (a "recall state") without exploding into runaway activity. This provides a robust mechanism for working memory even while synapses remain plastic.
- Associative Memory Recall: The network demonstrates pattern completion. Presenting only a partial stimulus can trigger the full assembly to "ignite," retrieving the complete memory trace.
- Biological Realism: Unlike many rigid models, this network reproduces the high trial-by-trial variability in firing observed in real cortical circuits.
Conclusion
By balancing Hebbian and non-Hebbian mechanisms, the authors demonstrate that spiking networks can achieve both rapid learning and long-term stability. This "orchestration" prevents catastrophic interference and runaway activity, offering a plausible framework for how biological brains manage the complex demands of memory formation and retrieval. Future work will explore how these rules scale to larger, more complex networks and the potential role of neuromodulators in gating these plastic processes.
Summary updated