Abstract
Neurons are the primary computational engines of the brain. Recent neurophysiological discoveries reveal that interneurons can slowly integrate spiking activity, sharing the output across a coupled network of axons to produce persistent firing even in the absence of somatic or dendritic input. This phenomenon is critical for understanding human cognition and working memory, yet conventional models fail to simulate it efficiently. We propose a computationally efficient neuron model that treats the axon as a slow leaky integrator, capturing nearly all known neural behaviors. This model controls the switching of axonal firing dynamics between passive conduction and persistent firing modes based on axonal integrated potential. We also demonstrate a persistent firing polychronous spiking network with asynchronous dynamics, proving the model's suitability for large-scale simulations and neuromorphic hardware design.
Introduction: The Axon as a Computational Unit
In classical neuroscience, the axon has traditionally been viewed as a passive "cable" responsible for the simple transmission of electrical spikes from the soma to other neurons. Popular efficient models, such as the Izhikevich model, focus almost exclusively on fast integration within the soma and dendrites over millisecond timescales.
However, recent evidence suggests that axons in certain interneurons possess independent computational capabilities. These axons can integrate spikes over seconds or minutes, eventually triggering "persistent firing"—a state where the neuron continues to pulse without further external stimulation. This mechanism is a likely candidate for the biological basis of working memory. Biologically detailed models like Hodgkin-Huxley can replicate this but are too computationally intensive for large-scale neuromorphic systems. There is a clear need for a model that is both biologically plausible and computationally efficient.
The Axonal Leaky Integrator Model
The proposed model extends the Izhikevich framework by introducing a third state variable, $w$, representing the "hyper-theoretical potential" of the axon. This addition allows the model to separate fast spike generation from slow axonal integration.
Dual-Threshold Mode Switching
The variable $w$ acts as a slow leaky integrator: it increases slightly with each somatic spike and decays slowly during periods of inactivity. The model employs a dual-threshold system to manage state transitions:
- Passive Conduction Mode: The default state where the axon simply transmits signals.
- Persistent Firing Mode: Triggered when $w$ reaches a high threshold ($w_p$). In this mode, model parameters are adjusted to transform the neuron into a self-sustaining oscillator.
- Recovery: Once $w$ leaks back below a lower threshold ($w_n$), the neuron returns to its passive state.
This architecture allows for a clear geometric interpretation of memory as a transition from a stable equilibrium to a limit cycle.
Experimental Validation and Key Results
The model was validated using biological stimulation protocols derived from rodent hippocampal interneuron experiments (Sheffield et al., 2011). Using current injections like "Step/Pause" and "Long Pulse," the researchers compared the model's output to real-world electrophysiology.
Biological Fidelity and Frequency Modulation
The model successfully replicated the characteristic 20mV drop in resting potential observed during persistent firing in biological neurons. Furthermore, it accurately simulated "firing frequency adaptation," where the rate of persistent firing gradually slows as the axonal potential $w$ decays.
Hardware and Network Scalability
Beyond mathematical simulation, the model was translated into an analog circuit design using SPICE and Verilog-A. The circuit, utilizing a parallel RC network and a Schmitt trigger, behaved identically to the mathematical model, confirming its viability for low-power neuromorphic chips.
In a large-scale simulation of 1,000 neurons with Spike-Timing-Dependent Plasticity (STDP), the inclusion of persistent firing interneurons led to asynchronous dynamics and reciprocal inhibition. These dynamics contribute to the beta and gamma oscillations observed in living brains, suggesting the model is robust enough for complex "polychronous" spiking networks capable of high-capacity pattern storage.
Discussion and Future Directions
By treating the axon as an independent computational unit, this model bridges the gap between simple mathematical neurons and complex biological reality. While currently a qualitative match for biological data, it provides a blueprint for artificial cognitive memory systems.
Future work will focus on utilizing these persistent firing networks to perform high-level cognitive tasks, such as sequence learning and pattern recognition. The ultimate goal is to develop bio-inspired computer architectures that manage memory with the efficiency and flexibility of the human brain, moving beyond the limitations of standard RAM-based systems.
Summary updated