🧬 biology

Single-Entity Spiking Neuron Models: Survey

This paper surveys and classifies various mathematical models for biologically plausible neural systems, encompassing single-entity spiking neurons alongside discrete and continuous analogs to accurately simulate membrane potential dynamics and other neural components.

Original authors: Leon Parepko, Danila Shulepin, Albert Nasybullin

Published 2026-07-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Leon Parepko, Danila Shulepin, Albert Nasybullin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine the human brain as a massive, bustling city. For a long time, scientists trying to understand how this city works have been building maps. Some maps are very rough sketches, while others are incredibly detailed blueprints. This paper is a guidebook that reviews different types of these "maps" (mathematical models) used to simulate how a single neuron (a single city block or building) behaves.

The authors, researchers from Innopolis University, focus specifically on Single-Entity Models. Think of these as models that treat a neuron as one single, indivisible room. They aren't trying to map out every tiny hallway, wire, and door inside the room (which would be a "multi-compartmental" model); they just want to understand how the room as a whole reacts to visitors.

Here is a breakdown of the different "architectural styles" they reviewed, using simple analogies:

1. The "Bucket" Models (Integrate and Fire)

These are the simplest maps. Imagine a neuron is just a bucket.

How it works: You pour water (input signals) into the bucket. As long as the water level stays low, nothing happens. But the moment the water hits a specific line (the threshold), the bucket instantly dumps its contents (fires a spike) and resets to empty.
The Catch: It's very fast and easy to build (like a plastic bucket), but it's not very realistic. Real neurons don't just "dump" water; they have complex chemical reactions. These models ignore the messy details of how the water actually moves inside the bucket. They are great for quick simulations but bad for studying the biology of the neuron itself.

2. The "Leaky Bucket" and "Smart Buckets" (LIF and Variations)

Scientists realized a real bucket isn't perfect; it has a hole in the bottom.

Leaky Integrate and Fire (LIF): This model adds a small hole to the bucket. If you stop pouring water, the bucket slowly drains. This is a bit more realistic.
The "Adaptive" Bucket: Some of these buckets get tired. After they dump water, they need a rest period before they can fill up again. Others change the "spill line" (threshold) depending on how much water they've handled recently.
The "Quadratic" Bucket: This one is even smarter. It can change its behavior based on how hard you push it, allowing it to handle complex patterns of input.

3. The "Detailed Blueprint" (Hodgkin-Huxley Models)

If the bucket models are plastic toys, the Hodgkin-Huxley (HH) model is a full engineering schematic of a real water treatment plant.

How it works: Instead of just "water," this model tracks specific chemicals (ions like sodium and potassium) moving through tiny gates (channels). It uses complex math to describe exactly how these gates open and close.
The Pros & Cons: It is incredibly accurate and looks just like a real biological neuron. However, it is so heavy and complex that running a simulation of a whole city (the brain) with these models would take forever. It's like trying to drive a city bus made of solid gold; it works perfectly, but it's too heavy to move.

4. The "Hybrid" Models (FitzHugh-Nagumo & Izhikevich)

To solve the "too heavy" problem, scientists created shortcuts that keep the best parts of the detailed blueprint but throw away the unnecessary weight.

FitzHugh-Nagumo: This simplifies the complex chemical gates into just two main variables. It's like saying, "We don't need to track every single ion; let's just track the 'fast' and 'slow' currents."
Izhikevich (IZ): This is the paper's "star student." It's a two-equation model that is as fast as the simple bucket but can mimic almost all the complex behaviors of the heavy gold bus. It's efficient enough to simulate thousands of neurons at once without breaking a sweat, making it a favorite for large-scale brain simulations.

5. The "Tree" Models (Dendritic Neuron Models)

So far, we've treated the neuron as a single room. But real neurons have branches (dendrites) that catch signals.

The DNM: This model treats the neuron like a tree with many branches. It doesn't just sum up all the rain hitting the tree; it looks at how the rain interacts on specific branches. It can "prune" (cut off) branches that aren't useful, creating a unique shape for specific tasks.
The Catch: While it handles the shape of the neuron well, it doesn't really explain the dynamic movement of electricity over time, and it can get very complicated if the tree has too many branches.

6. The "AI Mimic" (Artificial Neural Networks)

Finally, the authors looked at using modern AI (like CNN-LSTM) to learn how a neuron behaves.

How it works: Instead of writing equations to describe the neuron, you feed a computer program thousands of examples of a real neuron firing. The program learns the pattern and can then guess what the neuron will do next.
The Pros: It's incredibly versatile and can mimic weird, unusual neuron behaviors that standard equations can't handle. It's also very fast on modern computer chips (GPUs).
The Cons: It's a "black box." You don't know why it made a decision, only that it did. Also, it needs a lot of training time before it's useful, and if you ask it a question it hasn't seen before (outside its training), it might fail.

The Bottom Line

The paper concludes that there is no single "perfect" model.

If you need speed and are simulating a huge network, use the Izhikevich or Integrate-and-Fire models.
If you need biological accuracy to study how chemicals move, use the Hodgkin-Huxley model.
If you need to mimic complex, weird behaviors without understanding the biology, use Artificial Neural Networks.

The authors provide a "cheat sheet" (Table I) comparing all these models based on how many math variables they need, how much computer power they require, and how well they match real biological experiments. Their goal is to help researchers pick the right tool for the job without getting lost in the complexity.

Technical Summary: Single-Entity Spiking Neuron Models

Problem Statement
The accurate simulation of biological neural networks remains a complex challenge. While first and second-generation artificial neural networks (ANNs) utilize activation functions that fail to capture time-dependent voltage dynamics, Spiking Neural Networks (SNNs) offer a more biologically plausible alternative. However, a significant barrier to the effective deployment of SNNs is the lack of robust classification methods to distinguish between various mathematical models and select the most appropriate one for specific use cases. Existing literature often lacks a multifaceted survey that explicitly compares model features and analyzes the correlated problems they solve.

Methodology
This work presents a comprehensive survey and classification of mathematical models for single-entity (single-compartmental) neural systems. The authors categorize these models into two primary sets: single-entity models (where the cell is treated as a unified mathematical unit) and composite models (multi-compartmental), focusing exclusively on the former. The review evaluates models based on their prevalence, innovative perspectives, and ability to simulate biological processes, including membrane potential dynamics. The analysis covers a spectrum from biophysically meaningless abstractions to complex, biophysically plausible systems, comparing them across dimensions such as computational complexity, number of parameters, and the range of dynamical behaviors they can reproduce.

Key Contributions and Model Analysis
The paper systematically reviews and classifies several major categories of neuron models:

Integrate and Fire (IF) Models:
- Described as biophysically meaningless but computationally efficient, these models rely on ordinary differential equations (ODEs) to sum inputs and emit a spike upon reaching a threshold ( $V_{th}$ ).
- Variations include the Leaky-IF (LIF), which introduces membrane leakage; Adaptive LIF, which incorporates refractory periods and threshold adaptation; and Quadratic LIF, which supports bistable states and meta-dynamic behaviors (MDNs).
- Limitations: While suitable for neuromorphic hardware implementation due to simplicity, they fail to represent action potential duration, amplitude, and shape, making them unsuitable for detailed biological simulations.
Hodgkin-Huxley (HH) Based Models:
- The HH model is the standard for biophysical plausibility, utilizing four ODEs to model ion channel currents (sodium and potassium) via probabilistic kinetic and Markov modeling.
- Limitations: High computational cost and analytical complexity hinder large-scale simulations.
- Simplifications: The paper reviews reduced-order models derived from HH, including:
  - FitzHugh-Nagumo (FHN): A two-dimensional phenomenological simplification separating fast and slow currents.
  - Izhikevich (IZ) Model: A computationally efficient two-ODE model derived from bifurcation analysis. It balances biological realism with speed, capable of reproducing diverse spiking patterns (e.g., regular spiking, bursting) and is widely used in large-scale cortical simulations.
  - Other variants include Morris-Lecar, Hindmarsh-Rose, and Wilson models.
Dendritic Neuron Models (DNM):
- Unlike traditional weighted-sum approaches, DNMs account for dendritic shape and local synaptic interactions. They utilize a four-layer structure (synaptic, dendritic, membrane, cell) with sigmoidal processing.
- Capabilities: DNMs can prune unnecessary synapses to create task-specific structures.
- Limitations: They do not describe dynamic system aspects and suffer from the curse of dimensionality.
Artificial Neural Network (ANN) Approaches:
- The paper examines using second-generation ANNs, specifically CNN-LSTM architectures, to simulate single-neuron behavior.
- Findings: CNN-LSTMs showed superior performance in replicating membrane potential and ion channel conductances compared to traditional ODE approaches in certain contexts.
- Hybrid Approach: A custom layer of IZ neurons was proposed to address the CNN-LSTM's tendency to ignore long-duration firing patterns, requiring only a marginal increase in trainable parameters.
- Limitations: ANNs struggle with out-of-domain data and incur significant time costs during the training phase prior to simulation.

Results and Comparative Analysis
The study culminates in Table I, a comparative matrix evaluating models based on:

Continuity: Whether the model is discrete or continuous.
Classification: Whether the model is HH-based, Phenomenological (P), or IF-based (IFB).
Complexity: The number of variables, parameters, and Floating Point Operations (FLOPs).
Dynamical Capabilities: The specific firing patterns the model can reproduce (e.g., Regular Spiking, Tonic Burst, Low-threshold Spiking, Resonator).
Experimental Evidence: Whether the model has been validated against experimental data.

Key comparative findings include:

HH models offer high accuracy and experimental validation but at a high computational cost (~1200 FLOPs).
Izhikevich models provide a strong balance, offering "Any" dynamic behavior with significantly lower computational cost (~13 FLOPs) and high experimental validation.
IF-based models are the most computationally efficient (1–26 FLOPs) but generally lack experimental validation for complex dynamics and cannot reproduce all firing types.
ANN-based models (CNN-LSTM) possess the highest parameter count (~1.95 million) but offer versatility in mimicking atypical neural behaviors, though they face challenges with generalization to out-of-domain data.

Significance and Claims
The paper positions itself as a guideline for selecting appropriate modeling approaches for single-entity neural systems. Its primary significance lies in providing a structured classification that highlights the trade-offs between biological plausibility, computational efficiency, and dynamical richness. The authors claim that by categorizing models based on their benefits and limitations, researchers can better navigate the lack of robust classification methods in the field. The work emphasizes that while novel approaches (like hybrid ANN-IZ or DNM) offer specific advantages, no single model is universally superior; the choice depends on the specific requirements of the simulation, such as the need for biophysical detail versus large-scale scalability. The authors explicitly state that future work will extend this classification to composite (multi-compartmental) models.