Spike-based Learning Rules and Stabilization of Persistent Neural Activity

Xiaohui Xie, H. Sebastian SeungNeurIPS 1999

In this work, Xiaohui Xie and H. Sebastian Seung provide a mathematical bridge between the timing of individual neural spikes and the high-level stability of neural networks. They demonstrate how specific biological learning rules—where the order of spikes determines whether a connection strengthens or weakens—can be simplified into "rate-based" rules that depend on the derivative of a neuron's firing rate. Crucially, the authors show that a specific "anti-Hebbian" version of this rule naturally tunes a network to maintain persistent activity, effectively allowing it to store short-term memories without drifting.

The Fine-Tuning Problem in Short-Term Memory

In neuroscience, "persistent activity" refers to the ability of a group of neurons to keep firing at a specific rate even after a stimulus is gone. This phenomenon is widely considered the biological basis for short-term (working) memory. However, creating a computer model that replicates this is notoriously difficult. These models, often called neural integrators, require extreme "fine-tuning": if the synaptic connections are even slightly too strong, the activity explodes; if they are too weak, the memory fades away (drifts).

The gap the authors address is how the brain might achieve this precise tuning automatically. While most models use simplified "firing rates," real neurons communicate via discrete spikes. The authors investigate whether biological spike-based learning rules can provide a self-tuning mechanism to stabilize these memories.

Methods and Modeling

The researchers used a combination of mathematical derivation and numerical simulation to test their hypothesis:

  1. Mathematical Approximation: They started with a spike-based learning rule where the change in synaptic weight depends on the time difference between a presynaptic spike and a postsynaptic spike (the "pairing function"). By applying a Taylor expansion, they showed that if the pairing function is balanced (the amount of potentiation equals the amount of depression), the learning rule simplifies to one that depends on the time derivative of the firing rate ($\dot{\nu}$).
  2. Spiking Autapse Model: To test this, they simulated a "memory neuron" with an autapse (a synapse that loops back onto itself). They used an integrate-and-fire model where the neuron receives inputs from "burst" neurons (to set a memory) and a "tonic" neuron (to provide background activity).
  3. Learning Implementation: They applied the "differential anti-Hebbian" rule to the simulated neuron, allowing the synaptic weights to update based on the timing of spikes during intervals between bursts.

To validate the mathematical derivations, Poisson processes were used to generate synthetic pre- and post-synaptic spike trains. The pairing functions used in the models were inspired by real-world electrophysiological measurements from cortical and hippocampal synapses.

Key Insights

  • The Derivative is Key: The most significant insight is that if a spike-timing-dependent plasticity (STDP) curve has an equal area of potentiation ("up") and depression ("down"), the synapse doesn't care about the absolute firing rate. It only cares about how the rate is changing.
  • Anti-Hebbian Learning Equals Stability: In standard Hebbian learning ("cells that fire together, wire together"), a change in rate leads to a positive feedback loop that increases drift. In differential anti-Hebbian learning, the rule acts as negative feedback. If the firing rate starts to drift upward, the rule weakens the synapse; if it drifts downward, it strengthens it.
  • Self-Tuning Integrators: This provides a biological solution to the fine-tuning problem. Instead of a programmer manually setting synaptic weights to a perfect 1.000, the differential anti-Hebbian rule allows the network to "feel" the drift and correct its own weights until the drift is zero.

Results

The authors successfully derived the rule $W_{ij} \propto -\dot{\nu}_i \nu_j$ for differential anti-Hebbian learning. When applied to their models, this rule yielded two major outcomes:

1. Persistence Achieved In simulations, an "untuned" neuron failed to hold a memory, with its firing rate dropping immediately after an input burst. However, after applying the spike-based learning rule, the same neuron successfully maintained a steady firing rate for seconds.

Comparison of an untuned neuron failing to hold memory versus a tuned neuron maintaining persistent activity.

2. Superiority to Fixed Weights Interestingly, the authors found that a network that continues to learn is more stable than one with "optimally" fixed weights. Continuous fine-tuning reduced the residual drift even further by dynamically adapting to the specific firing rate at hand.

Continuous fine-tuning reduces residual drift compared to optimally fixed weights.

Limitations and Future Work

While promising, the model has a few limitations. First, the derivation assumes the pairing function's area is exactly zero (perfectly balanced). If there is a slight imbalance, the network will eventually drift, limiting how long a memory can be held. Second, the simulation focused on a single-neuron autapse. While the math suggests it should work for large recurrent networks (like the oculomotor system), the paper does not explicitly simulate a high-dimensional network.

Finally, the paper shows how to hold a state, but not necessarily how to compute or integrate new inputs (like adding vestibular signals to a head-direction memory). Finding rules for training networks to perform active integration remains an open challenge for future research.

Summary updated

BibTeX
@inproceedings{xie1999spike,
  author = {Xiaohui Xie and H. Sebastian Seung},
  title = {{Spike-based Learning Rules and Stabilization of Persistent Neural Activity}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1999},
}
Figure