Eugene Izhikevich and why the choice of spiking neuron model matters for consciousness

In computational neuroscience and neuromorphic engineering, the default model of a spiking neuron is the leaky integrate-and-fire (LIF) unit. It is computationally cheap, requiring only a single differential equation tracking membrane potential, and easy to implement in digital silicon. It is the model behind first-generation neuromorphic hardware like Intel’s original Loihi chip and the baseline architecture running in Layer 1 of the Substrate Console.

In two foundational papers published in IEEE Transactions on Neural Networks, “Simple model of spiking neurons” (2003, DOI: 10.1109/TNN.2003.820440) and “Which model to use for cortical spiking neurons?” (2004, DOI: 10.1109/TNN.2004.832719), applied mathematician Eugene M. Izhikevich proved that this simplicity comes at a steep theoretical cost. A standard LIF neuron is a one-dimensional dynamical system. It can integrate incoming current and discharge when a threshold is crossed, but it is mathematically incapable of reproducing the rich dynamical repertoire of biological cortical neurons, including bursting, chattering, spike-frequency adaptation, subthreshold oscillations, and post-inhibitory rebound.

For artificial intelligence tasks like static image classification, that omission is harmless. For consciousness science, it is decisive. Leading neurobiological theories of consciousness do not merely require that a network emits spikes. They require that the network produces specific, stable collective dynamics, such as self-sustaining reverberant gamma-band oscillations in Victor Lamme’s Recurrent Processing Theory (RPT) or non-linear workspace ignition in Stanislas Dehaene’s Global Neuronal Workspace (GNW). As Izhikevich demonstrated, those collective dynamics depend directly on the dynamical diversity of the underlying single-neuron models.

The Mathematical Difference Between LIF and Izhikevich

A standard leaky integrate-and-fire neuron models the membrane potential $v$ as a resistor-capacitor circuit. When input current arrives, $v$ leaks exponentially back toward resting potential or charges upward toward a fixed threshold $v_{th}$. Once $v \ge v_{th}$, an instantaneous spike is recorded, and $v$ is reset to $v_{reset}$. Because $v$ is the sole state variable, the neuron has no internal memory of its recent activity other than its current voltage. It cannot exhibit history-dependent behaviors like adaptation, where repeated stimulation progressively slows the firing rate, or resonance, where a neuron responds selectively to specific input frequencies.

Izhikevich’s 2003 model introduces a second dynamical variable, the membrane recovery variable $u$, which accounts for the activation of potassium ionic currents and inactivation of sodium ionic currents:

\(v' = 0.04v^2 + 5v + 140 - u + I\) \(u' = a(bv - u)\)

with the auxiliary after-spike resetting condition:

The four dimensionless parameters tune the cell’s dynamical bifurcation geometry:

• $a$ describes the time scale of the recovery variable $u$. Smaller values produce slower recovery.
• $b$ describes the sensitivity of $u$ to subthreshold fluctuations of $v$.
• $c$ describes the post-spike reset value of membrane potential $v$.
• $d$ describes post-spike reset of recovery $u$, modeling slow inward sodium and calcium currents.

By varying these four parameters, a single two-variable system can switch among regular spiking (RS) pyramidal cells, intrinsically bursting (IB) deep-layer cortical neurons, chattering (CH) 40 Hz gamma-generating cells, and fast-spiking (FS) GABAergic inhibitory interneurons.

Open the Substrate Console Layer 1 running in your browser. Load a region template built from the Allen, BrainGlobe or Julich-Brain atlases, change the thresholds and the connectivity, and watch leaky integrate and fire neurons spike.

Why Recurrent Processing Requires Dynamic Diversity

In Victor Lamme’s Recurrent Processing Theory, consciousness emerges when feedforward sensory signals trigger late, sustained reverberant feedback loops between cortical areas. In the visual cortex, this reverberation manifests as sustained oscillatory activity in the gamma band (30 to 80 Hz).

When researchers construct recurrent neural networks using homogeneous LIF units, sustaining this reverberant state is notoriously fragile. A recurrent network of identical LIF neurons typically suffers from a dynamical knife-edge: if synaptic recurrent weights are too low, reverberation immediately fades to silence; if weights are slightly increased, the network rapidly synchronizes into pathological, runaway epileptiform bursting.

Biological cortex avoids this runaway synchrony through cellular heterogeneity. Chattering (CH) neurons naturally pace gamma-frequency oscillations. Fast-spiking (FS) basket interneurons provide fast feedback inhibition that cuts off runaway excitation within milliseconds. Low-threshold spiking (LTS) interneurons provide slow dendritic inhibition that regulates long-term gain. As Izhikevich showed in his 2006 work on polychronization (Neural Computation, DOI: 10.1162/neco.2006.18.2.245), networks composed of heterogeneous Izhikevich neurons naturally support “polychronous groups”, reproducible, time-locked spiking sequences with millisecond precision that persist without requiring massive global synchronization.

Without the two-dimensional dynamical properties that Izhikevich formulated, a silicon network attempting to satisfy Lamme’s recurrent processing criterion must rely on artificial external tuning rather than self-stabilizing intrinsic dynamics.

Open the Origins Console An interactive map of where consciousness appears in the animal record. Feinberg and Mallatt place three separate origins across 560 million years, from the Cambrian to the present.

Implications for Global Workspace Ignition and IIT

The choice of neuron model is equally consequential for other leading consciousness theories:

1. Global Neuronal Workspace Theory (GNW): Stanislas Dehaene and Jean-Pierre Changeux emphasize that conscious access corresponds to an “all-or-none” ignition event, where local representations cross a threshold and broadcast globally across long-range cortico-cortical connections. In biological cortex, this ignition is mediated by thick-tufted Layer 5 pyramidal neurons, which exhibit intrinsic bursting and calcium-dependent dendritic spikes. An Izhikevich intrinsically bursting (IB) parameterization captures this non-linear burst threshold directly, whereas a standard 1D LIF neuron treats an ignition-inducing burst identically to a train of isolated spikes.

2. Integrated Information Theory (IIT): In IIT, as developed by Giulio Tononi and Christof Koch, the fundamental quantity $\Phi$ (phi) measures the intrinsic cause-effect power of a system’s causal structure. A system’s causal repertoire is derived from its transition probability matrix. A network of 1D memoryless LIF units has a vastly constrained state space compared to a network of 2D dynamical units where each node carries a continuous recovery state ($u$). The presence of internal recovery dynamics allows individual nodes to express history-dependent causal transitions, enriching the system’s overall cause-effect structure.

The Neuromorphic Shift from Loihi 1 to Loihi 2

The historical limitation preventing the widespread adoption of Izhikevich models in hardware was arithmetic complexity. While Izhikevich neurons require only 13 floating-point operations per millisecond step, compared to over 1,200 operations for a four-variable Hodgkin-Huxley conductance model, they still require quadratic terms and internal state tracking that first-generation fixed-function neuromorphic chips could not execute efficiently.

As examined in the analysis of Intel Loihi 2 and orbital neuromorphic computing, Mike Davies and the Intel Neuromorphic Computing Lab redesigned the core architecture between Loihi 1 and Loihi 2 to address this exact bottleneck. Loihi 2 replaced fixed LIF microcode with fully programmable neuron instruction sets. This architectural shift allows researchers to execute custom Izhikevich, Resonate-and-Fire, and adaptive threshold models directly on neuromorphic silicon.

For projects investigating emerging consciousness under functionalist emergentism, including The Consciousness AI (https://github.com/tlcdv/the_consciousness_ai), the lesson of Izhikevich’s work is clear: substrate independence cannot treat the neuron as a trivial point-process black box. If consciousness arises from the collective dynamical geometry of recurrent networks, the mathematical model of the single neuron must possess the dimensional degrees of freedom necessary to support those dynamics.

Eugene M. Izhikevich is CEO of Brain Corporation and author of Dynamical Systems in Neuroscience (MIT Press, 2007, ISBN: 978-0-262-09043-8). His landmark paper “Simple model of spiking neurons” appeared in IEEE Transactions on Neural Networks in 2003 (DOI: 10.1109/TNN.2003.820440).